fix: retry transient vLLM adapter reload failures

[yaodong-shen]

Summary

• retry LoRA reloads after connection establishment failures and transient 502/503/504 responses
• check the supervised vLLM child before every attempt so a real process exit stops retries immediately
• report the request URL, process state, runtime log path, checkpoint step/path, and final error when retries are exhausted

Fixes #678

Retry safety

The request is retried at most four times with exponential backoff. Client errors and read/write failures are not retried because the server may already have applied the non-idempotent adapter load.

Validation

• uv run --no-sync pytest -q tests/unit/test_vllm_runtime.py — 5 passed
• existing runtime launcher tests — 14 passed; one macOS /tmp versus /private/tmp assertion failed identically on a clean origin/main@3bdbf229 worktree
• prek ruff, ruff-format, ty, and uv-lock-check hooks — passed
• Python 3.12 compile and git diff --check — passed

The complete unit suite requires optional backend and GPU dependencies that are not installable in this macOS environment; repository CI installs those extras in its CUDA runner.

[yaodong-shen]

The Prek, Package Install, and TrainerRank GPU workflows are waiting for maintainer approval and created no jobs for this external-fork PR. The focused retry tests and the repository Prek checks pass locally. Could a maintainer approve the workflow runs when convenient?