[blackwell] sm100: FP8 blockwise GEMM -> torch._scaled_mm

[Andrewxu313]

Parent: #204 — Blackwell Phase 2 kernel port
Branch: tairan/blackwell-02-03-fp8-scaled-mm
Spec: batchgen-agent-metadata/batchgen_design/blackwell/blackwell-kernel-port-v1.md § Sub-task 3

What

Add arch == 'sm100' branch in batchgen/moe/grouped_fp8_blockwise_moe.py that calls torch._scaled_mm(a_fp8, b_fp8, scale_a=..., scale_b=..., out_dtype=bfloat16). Replaces the SM90a TMA+CuTe CUDA extension path on B200.

Acceptance

• GLM-5 single-layer FP8 MoE forward runs on B200
• max_rel_err < 1e-2 vs H20 baseline