vulkan: double the K-quant int-mmq large tile on RDNA3 iGPUs

[aic0d3r]

Implements #27553.

The integer-dot mmq path picks its large tile when both matmul dims exceed 64, which is every big GEMM in a dense-model prefill. On RDNA3 iGPUs (gfx1151-class, e.g. Radeon 8060S) the stock 128x128 tile at 256 threads underfills the shader cores at prefill shapes and re-reads the A weights once per 128-token slab; doubling only BN and the thread count keeps the per-thread register tiling identical while halving the A re-reads per token. No shader changes; wg_denoms kept in sync so dispatch and split-k see the real tile.

Qwen3.8-27B UD-Q5_K_XL-v2, RADV, -fa 1 -b/-ub 2048, same-binary stash A/B, r=5:

test	stock	128x256@512	gain
pp512	294.5 ±6.4	517.0 ±3.4	1.76x
pp4096	249.6 ±0.2	436.7 ±0.8	1.75x
pp16384	216.1 ±0.4	333.2 ±1.7	1.54x
tg128	10.76	10.82	unchanged

Also measured (r=3, pp4096): Q4_K_XL 465.1 ±0.8, Q6_K_XL 411.8 ±1.3 with the new tile.

Correctness and scope:

• Greedy outputs byte-identical (temp 0, diffed, -n 120).
• Decode and speculative-verify batches (n ≤ 64) select the small/medium tile and are structurally unaffected — verified with an MTP/DFlash2 spec-decode serving config: 28.7 t/s average decode at 64% draft acceptance, no regression vs baseline.
• Gated to AMD RDNA3 && device->uma; discrete GPUs, other vendors, and all mul_mat_id (MoE) tile families untouched.
• Neutral probes on this hardware: BM=256, 1024 threads, TM/TN/WMITER reshapings, wave32. The coopmat dequant family (iq4_xs path) measurably dislikes the doubled tile and stays stock.