Support max_kv_size configuration in HTTP server

[r-bahuguna]

Adds --max-kv-size as a server startup argument, bringing it to feature parity
with mlx_lm.generate().

Closes #615

Changes

• Added --max-kv-size CLI argument to mlx_lm.server
• Passed max_kv_size to make_prompt_cache() for both primary and draft models in _serve_single
• Passed max_kv_size into BatchGenerator for the batched request path
• Updated test mock to include the new max_kv_size default

Usage

python -m mlx_lm.server --model <model> --max-kv-size 4096