Disclaimer: A lot of this post was copied from my answer to a similar sort of game.
Summary
It turns out that this game is very closely related to the Octal Game with code "0.007". Using the Sprague-Grundy Theorem combined with a winning strategy for Nim, it is relatively straightforward to analyze small games like this (e.g. your game for small $n$), and there are theorems which give efficient algorithms to learn about this in the medium range. Some Octal games can be efficiently solved even in large cases, but this particular game is not one of them.
It turns out that $B$ has a winning strategy for $N={0, 1, 2, 8, 14, 24, 32, 34, 46, 56, 66, 78, 88, 100, 112, 120, 132, 134, 164, 172, 186, 196, 204, 284, 292, 304, 358, 1048, 2504, 2754, 2914, 3054, 3078, 7252, 7358, 7868, 16170}$ and $A$ has a winning strategy in all other cases up to $2^{28}-1$. But I believe the general result is still open.
Connecting this game to existing theory
First note that you shouldn't be able to draw a curve through points as then drawing a curve through all of the points would be a winning move all of the time. To connect this game to existing general theory, we can reconceptualize it in the way that Lærne did: After you draw a curve, you remove three points from a component, and split the remainder into two components (where one or both of those may be empty). That is traditionally encoded in a "number" with Octal digits. In this case, the code is $0.007$, where the $0$s indicate that you can't remove only one or two points, and the three bits of the $7$ indicate that you can leave zero or one or two components when you remove three points from a component.
Who wins?
Since the original question was "who has a winning strategy?", rather than "what is the winning strategy, I won't re-present an introduction to the relevant theorems in detail.
If you don't already know about the strategy for Nim or the Sprague-Grundy Theorem and how they can be applied to similar games, you can read one or more of the following:
Very briefly, positions in games like this act in combinations like single heaps of Nim, where the size of the corresponding Nim heap is the least number that's not a size of a Nim heap corresponding to a position you can move to. The strategy for Nim tells you that combining games with known equivalent Nim boils down to bitwise XOR (aka "adding in base $2$ without carrying").
Because of the strategy for Nim (how Grundy values combine), and a position in this game is a combination of separate components, it suffices to know the Grundy/Nim values for a single "heap" (in our context, a single component of points). There are tricks to make this calculation more efficient than the naive method, some of which are covered in the graduate textbook "Combinatorial Game Theory" by Aaron N. Siegel. For some similar games, the Nim values are known to be eventually periodic (by a theorem that says they will be if they look periodic long enough), but according to Flammenkamp, for $0.007$, $2^{28}$ values have been calculated with no periodic behavior, although the values $N$ for which $B$ wins appear to stop at $16170$.
Aaron Siegel's program Combinatorial Game Suite allows efficient calculation of middling values (but probably wouldn't get up to $2^{28}$. For example, the lines hr:=game.heap.HeapRules.MakeRules("0.007") and hr.NimValues(100000) takes about a minute and a half on my machine to produce the list of Nim values for $N=0$ to $N=99999$.
For example, the Nim values from $N=0$ to $N=499$ are as follows (remember that $N=0$ corresponds to a win for $B$) $0,0,0,1,1,1,2,2,0,3,3,1,1,1,0,4,3,3,3,2,2,2,4,4,0,5,5,2,2,2,3,3,0,5,0,1,1,1,3,3,3,5,6,4,4,1,0,5,5,6,6,2,7,7,7,8,0,1,9,2,7,2,3,3,3,9,0,5,4,4,8,6,6,2,7,1,1,1,0,5,5,9,3,1,8,2,8,5,0,1,1,12,2,2,7,3,3,9,4,4,0,11,3,3,3,9,2,2,8,1,3,5,0,9,12,2,6,13,13,5,0,1,1,4,11,7,7,10,3,4,1,4,0,5,0,3,3,6,7,2,14,13,10,4,12,9,2,2,3,3,6,9,9,1,16,4,8,3,3,2,15,1,1,4,0,5,5,16,6,6,6,8,0,16,5,4,4,17,2,2,7,14,6,10,12,1,0,16,13,3,6,2,7,7,8,1,0,5,17,2,12,15,3,11,0,19,18,12,4,16,17,2,2,21,6,9,4,19,5,5,17,10,3,6,19,2,7,8,4,1,9,12,7,2,13,6,3,19,5,9,4,8,8,17,17,2,15,18,1,1,8,5,21,16,21,3,19,19,13,5,18,1,4,17,7,2,7,6,3,19,12,5,5,16,16,6,17,19,7,7,18,1,4,17,0,9,16,3,3,14,13,22,0,1,15,24,17,2,6,18,3,4,19,19,0,8,21,16,3,15,7,26,18,13,1,1,17,9,2,21,2,6,22,19,9,5,16,4,16,20,3,7,18,23,22,8,20,5,16,21,15,6,10,19,18,18,18,4,4,17,17,7,2,3,23,19,9,5,0,16,16,3,17,30,2,18,18,8,4,17,17,9,27,6,10,19,19,14,9,9,4,20,17,14,11,7,18,6,19,19,5,13,16,16,10,6,19,19,23,18,4,4,17,12,12,14,10,6,3,19,5,9,5,21,16,20,6,7,7,18,30,13,13,17,12,21,15,10,3,19,22,18,8,4,32,17,17,11,14,6,26,24,12,5,9,16,16,6,7,7,7,18,18,8,4,17,20,7,16,10,10,22,19,22,9,23,4,13,17,20,7,11,23,23,4,5,9,5,16,16,10,17,10,22,18,23,8,4,17,17,20,16,32,13,19,19,33,5,5,24$
Best Answer
For a given combination of board size/rack size/dictionary size, there is obviously an optimal strategy for Scrabble, because the number of positions is finite. If the sequence of tiles is known in advance, then by "optimal strategy" I mean the strategy that maximises your final winning margin (or minimises your final losing margin); if the tiles are drawn at random, then you have to choose: do you want to maximise your expected winning margin, or maximise your chance of winning? But in any case the existence of a strategy is not in doubt.
But when you bring complexity into the picture, things get messy. There are too many parameters: board size, rack size, alphabet size, dictionary size, total number of tiles.
We might as well assume that the board is infinite in all directions, because the playing area is limited by the total number of tiles.
Now the big question is, are we using a fixed dictionary? If so, then the alphabet size is also fixed, and the only variables are rack size and total number of tiles. Then we have a simple polynomial-time algorithm for finding the highest-scoring move in a given position: for each starting square, and each direction (horizontal/vertical), and each word W in the dictionary, decide whether W is a valid move with the given rack; and choose the highest-scoring valid move. With a few obvious optimisations, this is not so very different from the actual algorithm used by the Scrabble program that I co-wrote 25 years ago, and it was lightning fast on a regular Scrabble board.
But this depends on the dictionary being in a nice format (e.g. a Directed Acyclic Word Graph). If the dictionary is allowed to vary, then I don't know if a polynomial-time algorithm is possible.
But in general, of course, the highest-scoring move is not always the best move. You must also try to maintain a playable rack, and avoid leaving high-scoring options for your opponent. The problem of finding the best move in this sense, even if the sequence of tiles is known in advance, is presumably NP-complete. Nevertheless, our Scrabble program was able to play the two-player endgame perfectly once the tile bag was empty!