ggml-webgpu: support non-square subgroup matrix configs for Intel GPUs

[SharmaRithik]

Overview

Enable WebGPU subgroup matrix support for Intel GPUs (Xe2/Battlemage).

Intel GPUs report non-square subgroup matrix configurations (e.g. M=8, N=16, K=16) via Dawn's ChromiumExperimentalSubgroupMatrix feature. The existing filter only accepted square configs (M==N==K), rejecting Intel GPUs entirely despite full hardware and driver support.

Changes:

• Relax the subgroup matrix config filter to accept any F16->F16 configuration, not just square ones
• Fix swapped subgroup matrix type parameters in flash_attn.wgsl (hidden by square configs where M==N==K)
• Add head dimension divisibility check in flash attention op support for non-square K/N dimensions

Tested on Intel Arc B580 (BMG/Xe2-HPG) with Mesa 25.2.8 and latest Dawn:

• MUL_MAT: 829/829 passed
• FLASH_ATTN_EXT: 2832/2832 passed

Additional information

Requires latest Dawn (April 2026+) which enables ChromiumExperimentalSubgroupMatrix for Intel Mesa drivers >= 25.2. Older Dawn versions do not expose this feature for Intel GPUs.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: no