Add KV cache quantization to mlx_lm.server (#1043)

[katlun-lgtm]

Add KV cache quantization to mlx_lm.server (closes #1043)

What

Exposes the KV-cache quantization already supported by generate/stream_generate
as server CLI flags:

• --kv-bits (default None → off)
• --kv-group-size (default 64)
• --quantized-kv-start (default 0)

When set, they are threaded into the server's sequential stream_generate path.

Why the batched path is handled by fallback, not plumbing

The server uses continuous batching (BatchGenerator) whenever a request
_is_batchable — which requires mergeable caches
(all(hasattr(c, "merge") ...)). QuantizedKVCache intentionally does not
implement merge(), so a quantized cache cannot participate in continuous
batching without new batch-membership support for quantized caches (a larger,
separate feature).

Rather than silently ignore --kv-bits in the default (batched) path, this PR
makes _is_batchable return False when --kv-bits is set, so those requests
use sequential generation, which applies the quantization. This keeps the
existing mergeable-cache invariant intact and makes the flag behave correctly.

Trade-off worth a maintainer call: with --kv-bits, requests no longer batch.
If you'd prefer, I can instead scope a follow-up that teaches the batched path
to carry quantized caches. Happy to go either way.

Changes

• mlx_lm/server.py: 3 argparse flags; pass to stream_generate; _is_batchable
  returns False when kv_bits is set.

Test plan

Verified on Apple Silicon (M3 Max, mlx 0.31.2, mlx-lm 0.31.3) with
mlx-community/Qwen2.5-0.5B-Instruct-4bit:

• --kv-bits flags parse with correct defaults (None / 64 / 0).
• stream_generate(..., kv_bits=4) produces a QuantizedKVCache and coherent
  output; kv_bits=None stays a plain KVCache (regression check).

kv_bits=None: cache=KVCache          quantized=False coherent=True
kv_bits=4:    cache=QuantizedKVCache quantized=True  coherent=True

• End-to-end HTTP smoke against a running mlx_lm.server --kv-bits 4 (recommended
  before merge; core generation path is exercised above).

Notes

Mirrors the existing generate.py flag names/help text verbatim for consistency.

[katlun-lgtm]

Verified end-to-end on M3 Max: mlx_lm.server --kv-bits 4 boots, /v1/chat/completions returns HTTP 200 and generates through the sequential (quantized-cache) path. The core stream_generate path is also confirmed to produce a QuantizedKVCache with kv_bits=4 (plain KVCache without). Output quality in any short snippet here reflects a 0.5B/4-bit test model, not the change itself.