feat(speco): add standalone TransferQueue (TQ) based draft feature transport and training

[xxyyrr598]

Summary

This PR introduces a standalone TQ (TransferQueue) training workflow for
draft-model feature transport and training. It decouples the drafter feature
producer and consumer through a queue-based protocol so that hidden states and
target logprobs are transferred off the driver, reducing driver-side memory
pressure and enabling independent, decoupled draft training.

Background / Motivation

Previously, draft-feature transport was tightly coupled to the training loop and
carried heavy SampleMetadata-based re-descriptions of the hidden-state layout,
which:

• prevented DraftFeatureSample.metadata from round-tripping cleanly;
• made the codec aware of DSpark/hidden-state layout internals;
• required touching the TQ protocol whenever a new algorithm was added;
• caused inconsistent stats between producer and consumer ready-tag filtering.

This PR reworks the transport layer around a single DraftFeatureSample
protocol and adds a standalone producer/consumer workflow.

Key Changes

Transport & protocol

• verl_speco/transport/drafter_sample_protocol.py: DraftFeatureSample codec
  with ready tags, EOS records and sample keys (schema v2).
• verl_speco/integration/transferqueue_bridge.py: TransferQueue bridge for
  restoring offloaded tensors on the consumer side.
• verl_speco/integration/mooncake_hidden_states_connector.py and
  verl_speco/trainer/mooncake_transfer.py: Mooncake-based hidden-state
  transfer.

Producer

• verl_speco/producer/input_reader.py: producer-side input reading.
• verl_speco/producer/vllm_feature_client.py: vLLM feature client pool with
  retries/backoff.
• verl_speco/standalone_tq_producer.py and verl_speco/tq_owner.py:
  standalone producer entry points and TQ ownership.

Consumer / training

• verl_speco/standalone_tq_training_launcher.py: standalone training launcher.
• verl_speco/trainer/tq_sample_source.py, trainer/tq_feature_store.py:
  queue-backed sample source and feature store.
• verl_speco/trainer/target_feature_replay.py,
  trainer/target_feature_pipeline.py: target feature replay/pipeline.
• verl_speco/trainer/standalone_checkpoint.py: standalone checkpoint export
  (including sharded lm_head export).
• Reworked trainer/draft_training_loop.py and trainer/feature_store.py to
  support the queue-based flow.

Scripts / examples

• tools/run_qwen3-8b_drafter_hidden_state_vllm.sh,
  tools/wait_for_vllm_endpoints.py.
• examples/run_qwen3-8b_drafter_dspark_separate_training.sh,
  examples/run_qwen3-8b_drafter_eagle3_separate_training.sh.

Docs & tests

• Design docs under docs/ (TQ protocol refactor, standalone producer /
  consumer / training, transferqueue integration plan).
• Unit tests for protocol, producer, TQ sample source, replay, transferqueue
  bridge and vLLM feature client.

Architecture

vLLM (hidden states)
        │  feature payloads
        ▼
 Producer (input_reader / vllm_feature_client)
        │  enqueue (TransferQueue / Mooncake)
        ▼
 Consumer (tq_sample_source / target_feature_replay)
        │
        ▼
 Draft trainer (draft_training_loop / feature_store)

Hidden states and large tensors are offloaded from the driver and restored on
the consumer via TQ keys, so only small ready tags cross the driver.

Testing

• New unit tests: tests/unit/test_* for the protocol, producer input reader,
  target feature replay, TQ consumer/producer, transferqueue bridge, vLLM
  feature client, and standalone launcher.
• pre-commit run --all-files passes (ruff, ruff-format, mypy, docs/license/
  naming sanity checks, compileall).

Notes

• This is an independent standalone training path, separate from the
  VeOmni-based co-training workflow.
• Requires the transferqueue package (see wheels/) and a vLLM instance
  serving hidden states with the target layer configuration.

Checklist

• pre-commit passes
• unit tests added
• CI verified