feat: vendor MXFP8 MoE runtime

[S1ro1]

Base of the NVFP4 stack: #9 → #12. Rebased onto current main; the package/registry implementation is unchanged. Indexed-attention PR #8 remains a separate child of this PR.

Summary

• Add a Python-only mxfp8_moe kernel package with differentiable grouped GEMM and explicit MXFP8 expert-parallel transport.
• Keep BF16 parameters and tensors at the public boundary; quantized data and scales remain internal.
• Vendor the orchestration from torchao commit 02105d46c with its BSD license and wide-MoE/large-tensor handling.
• Extend the kernel manifest, registry, and wheel builder to support Python-only kernels and runtime requirements.

Validation

• Built a wheel with PRIME_KERNELS=mxfp8_moe and PRIME_KERNELS_REQUIRE=1 and verified the module, manifest, and license are included.
• Imported the public module against the pinned torchao build.
• Ruff check, Ruff format check, and diff checks pass.
• The registry rejects the kernel on H200 before import as expected. Numerical execution still requires an SM100 runner.