Add alternating SVD refinement to SVDQuant (svd_refine_iters)

[Rudin6]

The one-shot SVDQuant split picks the low-rank branch from the smoothed weight's singular directions alone, so it cannot account for what the residual quantizer will get wrong. This adds an opt-in svd_refine_iters knob: each round refits the factors against

smoothed_weight - dequant(quantize(residual))

then re-anchors the residual on smoothed_weight, letting the low-rank branch absorb the residual quantizer's error pattern. Defaults to 0, which is the original one-shot behaviour.

The refinement lives in the new _refine_lowrank_split() in quantizer.py rather than inside decompose_lowrank_residual(), so the shared lowrank.py and packing.py stay untouched and svd_refine_iters=0 is a no-op by construction. Verified: with the knob at its default, the exported state_dict is bitwise identical to the pre-change output across 72 configurations (shape x precision x dtype x calibrate_precision x rank).

Two details the residual simulation has to get right, both pinned by tests against an independent replication of the packing path:

• The residual must be re-anchored in the dtype the SVD actually used (float64 on the "high" route, float32 otherwise, since bf16/fp16 SVD falls back to float32). math_dtype is the raw torch_dtype on the "low" route, and reconstructing in bf16 there drifts the next round's refit target.
• NVFP4 residual quantization is asymmetric: the packer normalizes by the torch_dtype group scales but pack_micro_scale() stores them as FP8 E4M3, so the kernel dequantizes with a different value than it normalized with. INT4 group scales are not micro-scales and stay at torch_dtype on both sides.

Averaged over 14 PixArt-Sigma attention and feed-forward layers with real calibration activations, 5 rounds cut weight error by 7.8-8.9% and layer output error by 17.4-18.2% versus the one-shot split, for both INT4 and NVFP4.

[DefTruth]

Thanks for your contribution! by the way, can you should some test results w/ or w/o this commit ?