[sglang-miles] Fix LoRA usage-counter lifecycle across cleanup paths

[Shi-Dong]

Problem

LoRARegistry.acquire() increments an in-flight usage counter before request
validation/dispatch, while /unload_lora_adapter waits for that counter to
return to exactly zero. Request cleanup paths that miss a release leave unload
blocked forever; duplicate releases can drive the counter below zero and have
the same effect.

The original version of this PR fixed the obvious abort and failed-dispatch
leaks by coupling rid_to_state.pop() with release(). That is not sufficient
for native parallel sampling (n > 1), where parent states, regenerated child
states, and a prefix-cache warm-up request all observe the same acquired LoRA
references.

Fix

This revision incorporates the lifecycle model from #31808 into the
sglang-miles branch and extends it for the request shapes used here:

• Add a per-ReqState exactly-once release guard. Normal completion, abort
  echo, scheduler error, and failed dispatch all route through the same helper.
• Keep strong references to asynchronous release tasks until they complete.
• Treat the acquired lora_id, rather than lora_path, as the ownership
  signal. This avoids release(None) when acquire never succeeded and remains
  compatible with [lora-related] feat(sglang-miles): Support upsert when loading adapters from tensors/distributed #30913, where an upsert deliberately reuses the stable
  adapter ID.
• For parallel sampling, make the prefix-cache warm-up state a non-owner and
  transfer the acquired references from bookkeeping parents to exactly
  batch_size * n regenerated child states.
• Support an explicit parent rid list: one parent state temporarily owns the
  n references for its prompt until fan-out, so validation/tokenization
  failure before fan-out still returns every acquired reference.
• Close the abort-vs-acquire race. If an abort removes a tokenizer-held state
  while registry acquisition is awaiting, the newly acquired orphaned ID is
  returned immediately.

The resulting invariant is:

Every successful LoRA acquire has exactly one owner and exactly one release,
across success, abort, validation failure, and native multi-LoRA parallel
sampling.

Relationship to earlier LoRA PRs

• [lora-related] feat(sglang-miles): Support upsert when loading adapters from tensors/distributed #30913 (LoRA upsert): already present in the sglang-miles base. This
  change releases the registry-provided stable lora_id, so an in-place upsert
  keeps the same identity and usage counter correctly.
• Fix LoRA usage-counter accounting across request lifecycle paths #31808 (usage-counter lifecycle): this PR is the sglang-miles port and
  superset of that lifecycle fix. The two PRs modify the same lifecycle lines,
  so they should not be independently stacked on sglang-miles; this revision
  already contains Fix LoRA usage-counter accounting across request lifecycle paths #31808's exactly-once guard, task retention, abort/error
  cleanup, and parallel warm-up ownership behavior. Shafeeq is retained as a
  co-author on the integration commit.

Tests

PYTHONPATH=python python3 -m pytest -q \
  test/registered/unit/managers/test_tokenizer_manager_rid_cleanup.py \
  test/registered/unit/managers/test_abort_request_prefix.py::TestAbortTokenizerHeldRequests \
  test/registered/unit/managers/test_io_struct.py

54 passed, 14 warnings in 11.74s

The lifecycle suite includes single and batch completion, abort/finish races,
acquire failure, abort racing acquire, n=4 prefix warm-up accounting, two
simultaneous LoRA adapters with explicit parent rids, and pre-fan-out failure.

Additional checks:

ruff check --ignore E402 <changed files>  # passed
python3 -m py_compile <changed files>     # passed
git diff --check                         # passed

---

CI States

Latest PR Test (Base): ❌ Run #31412296286
Latest PR Test (Extra): ❌ Run #31412295918
Latest PR Test (AMD ROCm 10): ➖ No AMD PR run found for this commit.