fix(ckpt): backport attention backend checkpoint inference to r0.5.0

[cuichenx]

Summary

• manually cherry-pick fix: honor attention backend for checkpoint inference #4370 into r0.5.0
• preserve --attention-backend in Megatron checkpoint inference overrides
• add release-compatible unit coverage for the checkpoint override helper

Conflict resolution

The auto cherry-pick conflicted because tests/unit_tests/scripts/test_text_generation.py does not exist on r0.5.0. I added a narrow release-compatible test file containing the #4370 attention-backend regression coverage, without main-only distributed-env tests.

Testing

• git diff --check --cached
• python3 -m py_compile scripts/inference/text_generation.py tests/unit_tests/scripts/test_text_generation.py
• uv run python -m pytest tests/unit_tests/scripts/test_text_generation.py -q (blocked locally: nvidia-resiliency-ext==0.6.0 has no compatible wheel for this host)
• uv run pre-commit run --all-files (blocked locally: same nvidia-resiliency-ext wheel/platform issue)

Cherry-pick of #4370.

[claude]

Light Code ReviewClean cherry-pick of #4370 to r0.5.0. The refactoring extracts _build_megatron_checkpoint_overrides and adds the missing attention_backend key to the Megatron checkpoint override dictionary. No logic errors or typos found.__Test coverage__The single test (test_megatron_checkpoint_overrides_preserve_attention_backend) covers the happy path where all optional fields are present. This is adequate for the targeted regression fix, but note that the negative paths are untested in this backport:- attention_backend=None (key should be absent from overrides)- Provider without cache_mla_latents attribute (key should be absent)- inference_moe_token_dispatcher_type=None (key should be absent)- dtype=torch.float16 (should produce bf16=False, fp16=True)These are low-risk since the conditionals are straightforward, and this is a release backport -- not the place to expand coverage.__Suggested test cases__No perf tests impacted.