fix(test): use SDPA for Nemotron-H HF reload

[yuhezhang-ai]

What

• select sdpa when checkpoint robustness loads a Nemotron-H model through Hugging Face remote code
• preserve flash_attention_2 for other remote-code models, including Nemotron-Flash
• preserve any explicitly configured attention implementation
• apply the selection to both source-load parity and post-checkpoint HF reloads
• keep nonzero ranks in a long-timeout synchronization while rank 0 performs the slow HF reload

Why

AM-672 fails while the vanilla-HF reference model runs through Hugging Face's FlashAttention varlen/unpadding path. Nemotron-H triggers a CUDA vectorized_gather_kernel index-out-of-bounds assert there. The later tensor-representation error and NCCL failures are secondary symptoms of rank 0's asynchronous CUDA failure.

PyTorch SDPA avoids the incompatible HF varlen path while retaining optimized CUDA attention dispatch. The explicit reload synchronization also prevents other ranks from hitting the default 600-second process-group timeout while rank 0 is still loading the 120B HF reference.

Validation

• python -m pytest tests/unit_tests/ci_tests/test_checkpoint_robustness_hf_kwargs.py -q — 6 passed, including the real two-process synchronization regression
• ruff format --check on both changed files
• ruff check on both changed files
• Scoped EOS/H100 release job passed: nemotron_super_v3_hellaswag_peft (job 362958206)
• Combined validation pipeline: https://gitlab-master.nvidia.com/dl/JoC/nemo-ci/-/pipelines/57979613

The separate vLLM deploy result in that combined pipeline belongs to #3061 / AM-668 and does not affect this HF checkpoint-robustness fix.

Linear: AM-672

[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 2ff349d

[yuhezhang-ai]

/ok to test 96423b4

[yuhezhang-ai]

@akoumpa can you bless this PR?