feat: add mlx chunked prefill support

[gufengc]

Summary

• Adds MLX/mac chunked prefill support with chunk-aware request length tracking and prefix-cache backed chunk progression.
• Makes prefix cache enabled by default and adds --chunked-prefill-size with a default of 1024; 0 disables chunking.
• Materializes MLX linear caches after prefill chunks to avoid lazy update graph accumulation across chunks.
• Preserves downstream pipeline chunk ordering when multiple chunks for the same request id arrive before the previous chunk finishes.

Why

Long-prefill requests on macOS MLX could exceed memory because the prefill path ran the full prompt at once. During chunked testing, downstream peers could also drop queued chunks with the same request id, causing two-node pipeline requests to hang. This PR chunks the MLX prefill work, reuses prefix cache state across chunks, and keeps distinct same-rid prefill chunks queued until their turn.

Validation

• pre-commit run --all-files
• pytest -> 143 passed, 5 skipped
• Manual macOS two-node validation with mlx-community/Qwen3.5-0.8B-MLX-bf16: node0 hosted layers [0, 12), node1 hosted layers [12, 24), --chunked-prefill-size 128; a 1675-token prompt completed successfully with 8 generated tokens.