feat(benchmark): per-model/GPU batch sizes and vocab projection for GEMM bench

[Z-Y00]

• Replace hardcoded BATCH_SIZE_LIST with dict aligned with primus config
• Add GPU_NAME_MAP to normalize GPU variants (MI355X -> MI35*, MI300X -> MI30*)
• Populate batch sizes from Primus training configs for MI35*/MI30* across dense models (Llama-2, Llama-3.1, Qwen2.5) and MoE models (DeepSeek-V2/V3, Grok-2, Mixtral, Qwen3)
• Add vocab_size to DenseModelConfigs and lm_head projection to gen_gemm_test_cases
• Update bench_gemm_turbo/torch/te to use get_batch_sizes(model, dtype, gpu)

[xiaobochen-amd]

Please rebase onto the latest main, then resubmit the code to trigger CI/CD.