fix(ci): honor checkpoint tokenizer in vLLM deploy

[yuhezhang-ai]

What does this PR do ?

Use the checkpoint-robustness tokenizer override when resolving arguments for vLLM deployment tests.

Changelog

• Resolve ci.checkpoint_robustness.tokenizer_name before falling back to the model path.
• Preserve an explicit --tokenizer as the highest-priority override.
• Add focused unit coverage for recipe and CLI tokenizer precedence.

Before your PR is "Ready for review"

Pre checks:

• Make sure you read and followed Contributor guidelines
• Did you write any new necessary tests?
• Did you add or update any necessary documentation? (No documentation change is needed for this resolver fix.)

Validation

• pytest tests/unit_tests/ci_tests/test_vllm_deploy_args.py -q (4 passed)
• Scoped checkpoint producer passed.
• Scoped vLLM deploy attempt and exact retry both passed tokenizer/model loading and completed generation, confirming the original NemotronHConfig / AutoTokenizer failure is fixed. Both subsequently reproduced the same existing five-token parity assertion (three matching tokens on prompt 2, on different EOS nodes); that follow-up is tracked in AMINT-285.

Additional Information

• Fixes AMINT-302.
• The post-fix numerical parity follow-up is tracked separately in AMINT-285.

[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.

[yuhezhang-ai]

/ok to test 1e5b2f5