Cherry pick weight_checker fp8 dequant fix and non-persistent buffer skip from #21494

[fzyzcjy]

Summary

Cherry-picks the weight_checker.py change from #21494 (originally landed on the sglang-miles branch) onto main.

Original author: @yym022502 (Yueming Yuan). I'm only the courier.

Changes

• Skip cos_sin_cache / freqs_cis buffers in _reset_tensors (non-persistent, get recomputed after weight load).
• Add weight_scale_inv names to skip_compare_names so fp8 scale tensors aren't double-compared.
• Make inverse_transform_scale_ue8m0 conditional on w_s.dtype == torch.int32 (UE8M0 packed format / Blackwell DeepGEMM only); other paths now feed w_s directly to block_quant_dequant.

PR chain

First of 3 PRs, all touching weight_checker.py only, cherry-picking weight_checker-related changes from sglang-miles to main:

1. This PR — cherry-pick of [sglang-miles] fix weight checker #21494
2. Cherry-pick of weight_checker portion of P2P Weight Update features for miles  #21278
3. Cherry-pick of weight_checker portion of [Hotfix] final fixes for P2P Transfer  #22663

The next two PRs each cherry-pick only the weight_checker.py slice of their source PR (the originals also touched many unrelated files). Please merge in order.

Test plan

• CI green

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!