feat(server): per-request ignore_eos via extra_body (vendor extension)

[Andrewxu313]

Description

Make ignore_eos a per-request vendor extension — passed via the OpenAI SDK's
extra_body (or a top-level key in a Batch JSONL body), the same convention vLLM/SGLang
use — and honor it per-sequence.

File	Change
server/io_struct.py	Add ignore_eos: Optional[bool] to ChatCompletionRequest + CompletionRequest
server/batch_scheduler.py	Pool mode: per-request req.body.ignore_eos with batch-level fallback, threaded via raw_request → admit entry; un-hardcode the legacy infer() path
batchgen_worker.py	Set seq.ignore_eos on admission; honor it per-sequence at the decode EOS boundary (eos_flags), OR'd with the global flag
worker/completion.py	is_sequence_completed / get_finish_reason OR in seq.ignore_eos (the completion gates)
sequence.py	Add ignore_eos to SequenceEntry.__slots__ + init (default False)
tests/worker/test_completion.py	Unit tests: per-sequence ignore_eos overrides EOS completion / finish_reason

Motivation

ignore_eos was only a batch-level CreateBatchRequest field, and pool mode (the
production path) never threaded it to the worker — the pool init message and
IntakeEntry.raw_request both omitted it, and the pool init handler never called
set_ignore_eos. So ignore_eos silently never held: sequences stopped at their real
EOS even when requested (observed: throughput_test.py → ignore_eos held: False, 707k of
8.39M expected tokens on a 2k→8k / 1024-seq run). It was also not a per-request field
and not an env var (BATCHGEN_IGNORE_EOS has zero readers).

The per-request extra_body form is the clean, extensible home for this and future vendor
flags (min_p, repetition_penalty, …). Backward compatible: the per-request value
falls back to the batch-level field, so this also fixes the previously-dropped batch-level
ignore_eos in pool mode — existing clients (incl. throughput_test.py) work unchanged.

Type of Change

• Bug fix
• New feature
• Breaking change
• Documentation update

Checklist

• I have read the CONTRIBUTION guide.
• I have updated the tests (if applicable).
• I have updated the documentation (if applicable).

Validation

• tests/worker/test_completion.py: per-sequence ignore_eos unit tests (CPU-only).
• Pending on H20: deploy + re-run the 2k→8k / 1024-seq throughput test → expect
  ignore_eos held: True and sum_completion_tokens == n*out_tokens.