Add an opt-in timeout for VLLMClient metadata requests

[030611]

What does this PR do?

Adds an optional metadata_timeout to VLLMClient for model discovery and world-size requests. A server that accepts either request but never responds now exhausts the existing read retries when a timeout is configured. The default remains unset.

This follows the metadata-only scope discussed in #6973 and leaves health probes (#6977), generation and weight-update timeout/retry policy separate. The value applies to connection establishment and read inactivity per attempt, rather than imposing an overall deadline.

Refs #6973. Discussion: #6973 (comment)

Local validation uses the complete production client and unchanged new test definitions with isolated collection. A localhost HTTP server exercises healthy and stalled responses, the existing retry count, and the separate POST timeout policy. GPU, NCCL and training integration were not run. Full test-file collection requires dependencies missing from this local environment.

Before submitting

• This PR fixes a typo or improves the docs (not applicable).
• Read the contributor guideline.
• Discussed the metadata slice in VLLMClient sets no request timeout, so a stalled vLLM server hangs training indefinitely #6973.
• Documented the new parameter and retry semantics.
• Added necessary regression tests.

AI writing disclosure

• No AI usage.
• AI-assisted: written and reviewed by a human.
• AI-generated: the PR was mostly or fully generated by an AI tool.

Codex prepared the implementation and tests. I reviewed the changes before submitting.

Validated on upstream 0b4b33d: 5 isolated-collection HTTP tests passed, and all changed-file pre-commit hooks passed. The full upstream test file and GPU/NCCL/training integrations remain unverified.

---

Note

Low Risk
Opt-in behavior with default unchanged; scope limited to two metadata GET paths and covered by new HTTP tests.

Overview
Adds an optional metadata_timeout on VLLMClient so model discovery (/v1/models) and get_world_size (/get_world_size) can fail instead of hanging when the server accepts but never responds. The value is passed as the requests timeout on those GETs only; default remains no timeout. Health checks, generation, and weight-update POSTs are unchanged.

Regression tests use a threaded localhost HTTP stub: healthy metadata with and without a timeout, stalled responses that exhaust read retries (six attempts), and confirmation that _post does not receive the metadata timeout.

^{Reviewed by Cursor Bugbot for commit a82eee0. Bugbot is set up for automated code reviews on this repo. Configure here.}