Support prefix cache hits in unified prefill path

[Kingwl]

This PR fixes a crash in the Metal unified paged prefill path when vLLM core prefix caching schedules a new request with num_computed_tokens > 0.

The root cause was that the unified path only tracked the current prefill chunk, but not the full request tokens, and it also assumed start_pos == 0 for complete prefills on new requests. This PR updates the model runner to carry both chunk tokens and full request tokens through the unified prefill flow, correctly updates paged sequence lengths for non-zero start_pos, and removes the incorrect assertion for new complete prefills with cached prefixes.

With this change, the long-context sonnet serve benchmark that previously failed with AssertionError: new complete prefill with start_pos > 0 not supported now completes successfully with prefix caching enabled.

Fixed #184

[WindChimeRan]

Thanks for the quick catch!

explicitly disable incompatible vLLM core prefix caching when this path is active.

I think we should explicitly disable prefix caching for now, because we have to and will support prefix caching on the paged path.

But prefix caching is not an one-day job. Please join the discussion in #182

Will take a closer look on your other fixes soon!