Add DeepSeek-V3 and DeepSeek-R1 on B200

[huydhn]

The main change is to add DeepSeek-V3 and DeepSeek-R1 on B200. To achieve this, I introduce a PLATFORM_SKIPS configuration to skip these models on A100 and H100. I'm also using this new configuration to shift Llama4 and bigger Gemma3 and Qwen3 to B200 while keeping Llama3 and smaller variants of these twos on A100/H100 (more coverage I guess).

For reference: configs for internal jobs for these models P1930074197

[huydhn]

The benchmark surfaces this recent error https://github.com/pytorch/pytorch-integration-testing/actions/runs/17631888148/job/50100699186#step:16:5520 from vLLM vllm-project/vllm#23582 (only on B200 I think). So, there are some missing metrics like latency or token/s for now. There is probably no action needed on the CI.