[CUDA] Ce fused kernel

[nastya236]

Fused backward + forward kernels for cross entropy.

There is an alternative to make this a custom kernel and put it into pretraining pipeline (as in mlx-lm for kl and js losses). But after given it a thought, I am leaning towards having a fast primitive for it. I feel that ce loss is an essential part of next token prediction and it deserves to be a fast primitive:)

forward: compute loss = logsumexp(x) - x_n -> logits and loss are saved for backward
backward: compute exp(loss + x_n) - onehot, donate logits buffer for dx (we can do it in most cases)

logits = model(x).astype(mx.float32) <- now after forward we don't need to cast to float32 and it saves crazy amount of memory: for 4B pretraining with input.shape = (2, 8192) and vocab = 128256 it results in more than 17GB decrease of peak memory!

I am happy to close this PR and make it a custom kernel if you think it is a preferable option.

[zcbenz]

The dropped peak memory is amazing!