fix: enable test_transform_bias_rescale_qkv_nested on XPU (issue #2182 fixed)

[laifenxiawucha]

The underlying C++ issue (missing T-rounding for nested tensors in
transform_bias_rescale_qkv_xpu) was fixed upstream via
intel/torch-xpu-ops#2450. This changes the unconditional XPU skip
to a no-op skip so the test runs now that the C++ fix is in place.

CUDA reference: aten/src/ATen/native/transformers/cuda/attention.cu:621-629

Reproducer

import math, torch
embed_dim, num_heads, bs, sl = 24, 4, 4, 2
x = torch.randn(bs, sl, 3*embed_dim)
xs = list(torch.unbind(x))
x_nt = torch.nested.nested_tensor(xs)
# The XPU kernel now correctly pads T to multiple of 8:
# Output q has shape [bs, num_heads, 8, dim_per_head]
# Before fix: output had T=2 (unpadded), reference expected T=8

Self-Review

• fix_location_correct — test skip removal only, C++ fix done in torch-xpu-ops#2450
• reference_semantics_matched — matches CUDA attention.cu:621-629
• minimal_change — 1-line skip change
• no_cuda_dependency
• test_coverage — tests 5 configs including non-multiple-of-8 sl
• blast_radius_checked — only affects the one test function
• no_unrelated_changes

Authored by Hermes Agent.