feat: add NVFP4 MoE kernels with optional 4over6 quantization

[S1ro1]

Add the SM100 nvfp4_moe kernel package for routed-expert training, exposed through the kernel registry and prime_rl::grouped_nvfp4_gemm. The implementation packs weights with per-expert global scales and activations with per-token global scales, runs NVFP4 grouped GEMM, and supports dequant_bf16 backward using the packed forward operands or bf16 backward using the master operands.

NVFP4GroupedGemm(four_over_six=True) enables adaptive 4/6 quantization for weights and both activation inputs. It matches FlashInfer's default 4/6 recipe: 448 normalization, MAE, strict reconstruction-error scoring, and default approximate candidate arithmetic. The option defaults to False. The wrapper remains traceable with torch.compile(fullgraph=False).

Declare the CUTLASS build dependency and header paths in kernels.toml, resolving build dependencies per selected kernel. Include the source license, package documentation, and focused numerical tests.

Stacked on #9 with gh stack: main → #9 → #12. This PR contains only the NVFP4 changes. The tested source commit 436719d and release wheel are unchanged. Prime-RL integration: PrimeIntellect-ai/prime-rl#3511.

Validation on GB200 with CUDA 13, torch 2.13.0 and FlashInfer 0.6.18, using the locally built release wheel:

• Eight GPU tests passed for 4/6 on/off, grouped forward, both backward policies, token locality, zero rows, and empty inputs.
• Fifteen comparisons against FlashInfer/vLLM default 4/6 matched packed values, logical block scales, and global scales exactly, covering contraction sizes 64–7168, irregular token groups, empty experts, and tiny blocks.
• Compiled Prime-RL backend forward and dequantized backward matched eager exactly. A configured MoE produced finite outputs and non-zero router and expert gradients.
• A two-step Qwen3-30B-A3B math RL smoke completed on 4 trainer + 4 inference GB200 GPUs, with the first 40/48 layers quantized and 4/6 enabled on both sides. Mismatch KL was 0.006612 and 0.005918; the previous 85% run measured 0.007310 and 0.006881. Sampled batches differ. Both smoke batches had zero task reward, so this validates execution and consistency, not learning quality.
• Lint, formatting, and whitespace checks passed. The wheel was built locally, without GitHub Actions.