[ci] Surface Anyscale logs for failed jobs

[bvolpato]

Summary

• Print a bounded remote log tail when an Anyscale job fails after its entrypoint starts.
• Keep log retrieval best-effort so diagnostics cannot mask the original job result.
• Allow the line count to be tuned with ANYSCALE_LOG_TAIL_LINES; default invalid values to 100 lines with a warning.

Why

Scheduled GPU failures currently stop at the Anyscale terminal state. Test output remains in Anyscale, so GitHub CI cannot distinguish a product regression from a remote environment failure. This adds failure context to the existing workflow log without changing successful jobs or capacity retry decisions.

Validation

• Mocked real-failure path: prints remote tail and exits 1.
• Mocked log-fetch failure: keeps original exit 1.
• Mocked capacity path: preserves retry behavior and does not print a misleading test tail.
• Mocked invalid line-count override: warns, falls back to 100, prints the failure tail, and exits 1.
• bash -n ci/submit_anyscale_job.sh
• ShellCheck
• gitleaks pre-commit hook
• uv lock --check
• git diff --check

[gemini-code-assist]

Code Review

This pull request adds best-effort remote log tail printing to the Anyscale job submission script when a job fails after its entrypoint has started. This is achieved by introducing a new print_remote_log_tail function and a configurable ANYSCALE_LOG_TAIL_LINES environment variable. The review feedback suggests validating that LOG_TAIL_LINES is a valid positive integer to prevent command failures, and appending || true to the print_remote_log_tail call to ensure that any unexpected failure during log retrieval does not interfere with the script's final exit status.