added b300/b200 support for perf plugins. B200/B300 must have expenda…

[rsalagame-nvidia]

Adding B300/B200 support for expendable segements. Resolved CUDA OOM issue as noticed from QA test cycle nvbug: https://nvbugspro.nvidia.com/bug/6160449

[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]

Light Review: The change is correct. One note: the deepseek_v3 block (line 299) is missing b300/gb300 in its gpu list despite having configs for those GPUs -- worth a follow-up. Suggested test cases: llama31_405b_pretrain_config_b300 (bf16, fp8_cs, fp8_mx, nvfp4), llama31_405b_pretrain_config_b200 (bf16, fp8_cs, fp8_mx, nvfp4). bf16 variants are negative controls (should NOT get expandable_segments set).

[claude]

Light Review: The change is correct. Potential issue (not blocking): deepseek_v3 block at line 299 does not include b300 or gb300 in its gpu list -- worth a follow-up check. Suggested test cases: llama31_405b_pretrain_config_b300 (bf16, fp8_cs, fp8_mx, nvfp4), llama31_405b_pretrain_config_b200 (bf16, fp8_cs, fp8_mx, nvfp4). bf16 is a negative control (should NOT get expandable_segments set).

[yaoyu-33]

@rsalagame-nvidia can you check ai comment + resolve conflicts.

[malay-nagda]

@rsalagame-nvidia this might not be needed after- #3781.
this flag is enabled for all cases except when FSDP/nccl_ub is used.