Improve Nemotron3 Super B200 BF16 Config

[zuriz-nv]

Update Nemotron3 Super B200 BF16 Config to use GB200 Config

• Logic remains the same for B200 FP8 and NVFP4 Configs.
• NEMOTRON_3_SUPER_PRETRAIN_CONFIG_B200_NVFP4_V1 already follows BASE_NEMOTRON_3_SUPER_CONFIG_GB200, meaning changing BASE_NEMOTRON_3_SUPER_CONFIG_B200 would allow BF16 and NVFP4 to directly use it, only needing to change NEMOTRON_3_SUPER_PRETRAIN_CONFIG_B200_FP8_MX_V1.
• BASE_NEMOTRON_3_SUPER_CONFIG_B200's expert_model_parallel_size=64 is removed as it is already in BASE_NEMOTRON_3_SUPER_CONFIG

[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.

[claude]

LGTM. The refactoring is correct. BF16 is intentionally changed to match the GB200 config (tp=2, CUDA graphs enabled, no recompute). FP8_MX and NVFP4 are preserved with identical effective configs. Removing expert_model_parallel_size=64 from BASE_B200 is correct since BASE_NEMOTRON_3_SUPER_CONFIG already sets it to 64. Suggested test cases: test_nemotron_3_super_perf_config_instantiation and test_nemotron_3_super_perf_config_nvfp4 (both existing, only exercise GB300, do not cover B200). No perf tests impacted.

[malay-nagda]

@zuriz-nv can you post the TFLOPs and step time you get for all 3 precisions with these changes?

[zuriz-nv]

Closing this PR as it has been superseded by PR #4621