feat: sync Nota XNNPACK with Google upstream

[ThibaultCastells]

Summary

• Rebase nota-github/XNNPACK onto current google/XNNPACK master at f45b3ea023012bd29deaa1900272933bca798ab5.
• Preserve Nota's Arm QD8/F16/QB4W GEMM config fix on top as f8d2ca0a2efbb9c97da9039933960748ebc6b2dc.
• Keep the fork one Nota commit ahead of current Google upstream, with no missing upstream commits.

Background

The previous Nota fork head was based on Google XNNPACK f0b3a2996d6fff82cd3f004c0b29f5ccb0ab947d and carried a custom fix for xnn_init_qd8_f16_qb4w_gemm_config().

Current Google master still gates that getter on qd8_f16_qb4w_gemm_config.arch, while init_qd8_f16_qb4w_gemm_config() does not assign .arch on its Arm branches. The original failure therefore remains present upstream: supported Arm hardware can receive a null GEMM config and fail runtime creation with xnn_status_unsupported_hardware.

The custom change is replayed after the upstream rebase and still returns the initialized config unconditionally, matching the qd8_f16_qc4w sibling behavior.

Test plan

• git rev-list --left-right --count upstream/master...HEAD -> 0 1
• git diff --check upstream/master...HEAD
• Confirmed current upstream still has the .arch return guard and no corresponding .arch assignment in the initializer.
• Confirmed the resulting diff against upstream is only the preserved one-line getter fix.
• Full platform validation is deferred to repository CI and the downstream ExecuTorch wheel/runtime matrix.