feat: add sglang chunked prefill support

[gufengc]

Summary

• Add minimal SGLang chunked-prefill glue that reuses native SGLang Req, PrefillAdder, ScheduleBatch, and unfinished chunk cache lifecycle.
• Align MLX middle-chunk behavior with SGLang semantics through a shared backend-neutral request-state helper.
• Pass chunked_prefill_size explicitly into the scheduler and stop reading chunk size indirectly from the cache manager.

Why

Parallax constructs ScheduleBatch directly instead of using the upstream SGLang scheduler, so GPU chunked prefill needed a thin adapter around SGLang's native chunk selection and lifecycle primitives. This keeps the SGLang changes small while avoiding divergent MLX/SGLang chunk semantics.

Validation

• uv run --no-sync pre-commit run --all-files
• uv run --no-sync pytest
• Manual SGLang single-node streaming test on my-windows with --chunked-prefill-size 128
• Manual Mac + Windows/H200 two-node pipeline test with layers split [0,14) and [14,28), validating long-prompt semantic output