feat(token-capture): add Megatron ledger capture backend

[lauradang]

Summary

• add deferred Megatron-Inference ledger capture and token-free pending call records
• extract vLLM worker and Megatron ledger capture into typed backend handlers with a shared lifecycle
• preserve served response IDs and token-chain custody through deferred MInf staging
• stage Megatron records at rollout completion while preserving existing vLLM worker staging behavior
• add ledger, adapter, handler, and model-server integration coverage

Testing

• pre-commit run --files on all changed working-tree files
• Gym token-capture and focused model-server suites — 268 passed

Baseline note

The complete responses_api_models/vllm_model/tests/test_app.py run has 15 response-ID expectation failures (chtcmpl-123 versus resp_123). A representative failure reproduces unchanged on pthombre/tq-tokidcap-capture, so it is not introduced by this branch.

Dependencies

• This draft intentionally targets pthombre/tq-tokidcap-capture, not main.
• RL consumer: feat(rollout): integrate MInf ledger capture with checkpointing RL#3869.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.