Add pre-quantized FP4 MoE weight loading

[bgchun-fs]

Description

Add support for loading pre-quantized FP4 MoE weights, skipping the runtime FP8→FP4 dequant→requant cycle during model startup.

Currently, FP8 MoE models go through dequant→FP32→requant→FP4 at load time, which takes ~45 min on CPU for DeepSeek-V3 671B.

Pre-quantizing MoE experts to FP4 offline reduces total model size from ~650 GB (full FP8) to ~338 GB (~48% reduction), since MoE expert weights make up the majority of the model.

This PR adds:

• MOE_SKIP_REQUANTIZE env var to skip runtime requantization
• create_weights override to allocate uint8-packed FP4 weight buffers
• uint8→float4_e2m1fn unpacking in process_weights_after_loading
• DSV3 converter script (scripts/convert/dsv3_converter.py) that converts FP8 2D-subchannel [128,128] → 1D-subchannel [1,N] with optional --fp4 flag for MoE expert FP4 packing

If the change fixes a Github issue, please include a link, e.g.,:
FIXES: #123456

Tests

Tested end-to-end with DeepSeek-V3.1 671B (vllm path, DP attention):

1. Converted weights with dsv3_converter.py --fp4
2. Served with MOE_SKIP_REQUANTIZE=1 MOE_REQUANTIZE_BLOCK_SIZE=512
3. Verified correct inference output

Checklist

Before submitting this PR, please make sure:

• I have performed a self-review of my code.
• I have necessary comments in my code, particularly in hard-to-understand areas.
• I have made or will make corresponding changes to any relevant documentation.

[bgchun-fs]

@jrplatin Hi, I tried to figure out how you compiled the following models. Could you verify whether this is correct? Thanks.

• jrplatin/DeepSeek-R1-1D-Subchannel-256
• jrplatin/DeepSeek-R1-1D-Subchannel-256-Packed