[TOP 2/N] Add batch-invariant scorer integration coverage

[kaixih]

Summary

Add one GPU integration test for the existing Miles true-on-policy scorer.

With Megatron batch-invariant mode enabled, the scorer keeps its existing path:

TP full-vocab gather
-> padded-vocab truncation
-> torch.log_softmax
-> selected-token gather

The test uses BF16 logits with the DSV4 129280-token vocabulary. It verifies
that Miles reaches Megatron's registered BI log_softmax implementation and
that the same row scored alone or in a 17-row batch produces a bitwise-identical
selected-token logprob.

Scope

• No runtime code or CLI changes.
• No duplicate BI kernel or custom op.
• DSV4 prefill-vs-trainer E2E parity remains a follow-up.

This is split 2/N from #1788 and is orthogonal to #1912.

Validation

• New GPU integration test: passed on GB200 and in stage-b-2-gpu-h200 CI.
• Focused local scorer/config tests: 59 passed.
• GitHub pre-commit and CPU matrix: passed.

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

[kaixih]

Closing this test-only split because no additional Miles runtime integration is required. Megatron-LM's enable_batch_invariant_mode() globally registers the accelerator implementation of aten::_log_softmax; therefore Miles' existing torch.log_softmax scorer automatically uses the batch-invariant kernel under --batch-invariant-mode. This conclusion is also recorded in #1788.