[Draft] Add PyTorch engine support for Hy3 BF16 and static FP8 (NO MTP)

[yidingcheng0206]

Summary

• Add PyTorch engine support for Hy3 BF16 inference.
• Add static per-tensor FP8 Linear and fused MoE support.
• Add Hy3 BF16 and FP8 checkpoint weight loading.
• Add TP-compatible packed QKV, dense MLP, shared MLP and routed expert loading.
• Add model construction, weight loading and static FP8 kernel tests.

Implementation

This PR is organized into two commits:

1. Add generic static per-tensor FP8 Linear and fused MoE infrastructure.
2. Add the Hy3 model implementation, module registration and model tests.

Validation

• ruff check: passed
• git diff --check: passed
• Hy3 tests: 9 passed
• Static FP8 fused MoE kernel tests: 2 passed
• Total: 11 passed

Serving validation:

• Hy3 BF16 TP8 serving passed on 8 NVIDIA H200 GPUs.
• Hy3 FP8 TP8 serving passed on 8 NVIDIA H200 GPUs.
• BF16 memory usage: approximately 77.5 GiB per GPU.
• FP8 memory usage: approximately 44.7 GiB per GPU.
• Non-MTP checkpoint weight mapping coverage was verified.

Limitations

• MTP is not supported yet.
• Static FP8 MoE currently supports TP but not EP.
• Full benchmark accuracy alignment is still in progress.
• The minimum required Transformers version needs further discussion.

Checklist

• BF16 model construction and weight loading
• Static FP8 Linear
• Static FP8 fused MoE
• TP8 serving smoke test
• Unit and kernel tests
• MTP support
• Full accuracy benchmark alignment

[grimoire]

if hy3 shares the same configuration behaviour as default, the config builder is not necessary.

[grimoire]

We have build_moe_all_reduce logical in moe module, read qwen3_5_moe.py as reference.