refactor: eliminate DisaggPDChunkedPrefillScheduler.

[weizhehuang0827]

Summary

• Eliminate DisaggPDChunkedPrefillScheduler class entirely
• PD + chunked_prefill now goes through DisaggPDScheduler → ContinuousScheduler::prepare_batch() → PrefillFirstPolicy
• Add full-footprint admission gate to prevent deadlock when multiple large requests compete for blocks
• Follows up on refactor: unify scheduler subclasses into ContinuousScheduler + SchedulerPolicy. #1992 (scheduler policy refactor)

Motivation

After #1992 unified scheduling logic into SchedulerPolicy, DisaggPDChunkedPrefillScheduler duplicated the same scheduling work in its own prepare_batch(). The key insight is that PrefillFirstPolicy already schedules chunk_queue (in-flight) before prefill_queue (fresh), which naturally prevents the hold-and-wait deadlock that the old class was designed to solve.

Changes

1. Delete disagg_pd_chunked_prefill_scheduler.cpp/.h and its test
2. Remove DisaggPDScheduler::prepare_batch() override — defer to ContinuousScheduler::prepare_batch() + policy
3. Remove DISAGG_PD_CHUNKED_PREFILL from SchedulerKind enum and factory
4. Add full-footprint admission gate in schedule_prefill_from_queue: fresh requests check if their total block footprint fits before admission (defense-in-depth against deadlock)
5. PD PREFILL forces exclusive batch via resolve_batch_mode (enable_mix_batch = false)
6. Skip expand_sequences for PD PREFILL in drain_request_queue (best_of_n deferred to decode)
7. Remove redundant external deallocate — allocate() internally handles cleanup on failure

Test Plan

• Unit tests: 42 pass (including new FullFootprintAdmissionGate and FullFootprintAdmitsBothWhenFits)
• PD + chunked_prefill=true + prefix_cache=false: shared-prefix 50 reqs, sharegpt 100 reqs, random 50 reqs — all PASS
• PD + chunked_prefill=false + prefix_cache=false: same tests — all PASS
• PD + prefix_cache=true: see bugfix: skip prefix-cached blocks during PD KV transfer. #2040 for a separate bugfix needed (pre-existing bug unrelated to this refactor)

Change Type

• Refactoring (no behavior change)

Reviewer Notes

The prefix cache + PD crash (batch_input_builder.cpp:269: remote block coverage shortage) is a pre-existing bug that also reproduces on ca6a5fd7e (before the scheduler refactor). Fix submitted separately in #2040.