[OOB][Perf] aten::native_batch_norm performance gap on B580 (13 models affected)

## Summary

`aten::native_batch_norm` is the top 4 op priority on B580 for XPU performance improvement in OOB 300 eager inference (fp16). The gap is significantly larger on B580 than G31, suggesting a platform-specific kernel issue.

## Data (from roofline profiling, 154 models, fp16 inference)

| Platform | Geomean R Delta | Models Affected | Total Saving (ms) |
|----------|------:|------:|------:|
| B580 | +0.0049 | 13 | 6,168 |
| G31 (B70) | +0.0012 | 10 | 145 |

Current fleet geomean R: B580=0.644, G31=0.675, 4080S=0.683.
## Note

The B580 total saving (6,168 ms) is 42x larger than G31 (145 ms) despite fewer affected models. This strongly suggests a B580-specific performance issue with the batch_norm kernel — likely related to occupancy, vectorization, or memory access patterns on BMG architecture.

Batch norm is a memory-bound op (vector engine). The roofline projects it at bytes/BW, so a gap means the kernel is not achieving peak bandwidth on B580.

## Hardware

- B580: 93 TFLOPS fp16, 410 GB/s
- G31 (B70): 154 TFLOPS fp16, 532 GB/s
- 4080S: 100.96 TFLOPS fp16, 716.8 GB/s

Start at the aten::native_batch_norm entry point and reproduce the fp16 eager-inference measurements on B580, comparing them with G31. Use the roofline data and the 13 affected models to determine whether the gap is reproducible; done means a measured B580 performance improvement with updated results.

Written by the indexing model from the issue text.