cuda: BATCH_INVARIANT: occupancy-independent FA KV split, PTQ1_0 mat-vec up to 8 columns

[professorpalmer]

One of the five pieces of #285, split per bri-prism's request on #317. Single commit on current prism (eaecb50c7), applies without conflicts. Only active under GGML_CUDA_BATCH_INVARIANT=1.

What it does

• The flash-attention vector path's KV split no longer depends on the occupancy of the kernel instance that runs, so the same query produces the same split whether it is decoded alone or inside a verify batch.
• PTQ1_0 batches of 5-8 columns stay on the PT mat-vec path instead of crossing to MMQ: a 5-column verify on MMQ did not match a single decode. GGML_CUDA_PTQ1_MMVQ_MAX overrides the crossover; the default of 4 is confirmed by measurement (pp5: 172 tok/s on MMQ vs 162 on the mat-vec, so 4 is where MMQ starts to win and where the verify columns of a draft of 4 still match).

Receipts

• test-backend-ops -b CUDA0 with GGML_CUDA_BATCH_INVARIANT=1: FLASH_ATTN_EXT 2994/2994, MUL_MAT 1516/1516.
• Keeps the weight path exact between drafted and undrafted decode. Attention is still not batch-invariant under drafting in general (its KV split follows the kernel instance and the padded KV length), which is stated in cuda+server: q8_0 K/V at the full 262k window on 12 GB (tiered VMM KV cache), MTP drafting at every depth #285 and in the Bonsai repo's docs.