split logsumexp

[awni]

Not sure it's worth merging this. The standalone benchmark is much improved but it's a very modest gain even for a very small model.

Adds a split logsumexp dispatch for when we get a really long vector.

x = mx.random.uniform(shape=(1, 4096 * 50))

def fun(x):
    for _ in range(100):
        x = x - mx.logsumexp(x, axis=-1, keepdims=True)
    return x

Pre: 234 ms
Post: 138 ms

Inference speed improved slightly for small Gemmas (which have a large vocab):

mlx_lm.generate --model mlx-community/gemma-3-1b-it-4bit --prompt "Write a story about Einstein" -m 512

Pre: 333.652 tokens-per-sec
Post: 334.837 tokens-per-sec

[zcbenz]

I'm closing this PR since the work will unlikely be continued.