[Refactor] T7: PreparedRequestQueue — parallel impl, consolidate with #903 (#661)

[XinhaoTheo]

Motivation

RFC #661 Template 7. The tri-state preprocessing→AR-engine handoff — prepared / inflight / aborted + a lock + the per-model preprocessing context — is duplicated line for line across moss_tts and moss_tts_local (~80 LOC each). Per #661 (framework owns reusable mechanics; model dirs own model semantics), this extracts that shared machinery into one generic queue and migrates both models onto it.

Note: #903 (@MelodyyyYin) implements the same template. happy to defer to #903 with a Co-authored-by credit, or fold my addition (below) into it.

Modifications

• New sglang_omni/scheduling/prepared_request_queue.py: generic PreparedRequestQueue[CtxT, PrepT] owning the tri-state + lock + the opaque per-model context. begin() reads the context and marks the request in-flight under one lock (preserves the original single-lock atomicity so a concurrent clear_context can't strand a stale in-flight id); publish() returns a non-observable bool; abort / pop / fail_inflight / set_context / clear_context.
• Migrated moss_tts and moss_tts_local request_builders.py to delegate to the queue; the duplicated per-model globals + lock are deleted in the same change. stages.py wiring (abort_callback / set-context) is unchanged.
• Updated both models' test_pipeline.py assertions to read _QUEUE.*.
• New CPU contract test: the seven §13 prepared/inflight/aborted scenarios plus a concurrency thread-storm.

Related Issues

Part of #661 (Template 7). Overlaps with #903 (same template) — see Motivation.

Accuracy Test

N/A — no model-side code (kernels / architecture / sampling) is touched. This is a behavior-preserving scheduling-layer refactor: the queue logic is line-for-line equivalent to the previous per-model code, and publish()'s bool return is non-observable (callers ignore it). Per-model Seed-TTS accuracy CI is the backstop.

Benchmark & Profiling

N/A — no performance-relevant change. Same lock, same operations per request; one
shared class replaces two copies.

Checklist

• Format your code according with pre-commit.
• Add unit tests.
• Update documentation / docstrings / example tutorials as needed.
• Provide throughput / latency benchmark results and accuracy evaluation results as needed.