HIP: tune MMVQ batch thresholds on RDNA3.5

[SimonTeixidor]

Overview

ggml_cuda_should_use_mmvq had no per-type tuning for RDNA3.5 and fell through
to a flat ne11 <= MMVQ_MAX_BATCH_SIZE (8). This PR adds a table for it which
gives a performance boost for batched generation.

Additional information

I introduced a temporary env var override which allowed me to sweep ne11 1-16 with
test-backend-ops perf -o MUL_MAT.

Results

threshold	types
ne11 <= 2	Q4_K, Q5_K
ne11 <= 3	Q3_K, Q6_K
ne11 <= 4	Q8_0, Q2_K
ne11 <= 5	Q4_1, Q5_1, MXFP4, IQ4_NL, IQ3_XXS
ne11 <= 6	Q4_0, Q5_0, IQ3_S, IQ4_XS, IQ1_S
default (8)	IQ2_XXS, IQ2_XS, IQ2_S, Q1_0, Q2_0, NVFP4

End-to-end (Strix Halo / gfx1151, Qwen3.8-27B-UD-IQ4_XS, llama-batched-bench, S_TG)

npl	gain
1	-0.8% (identical to master, null control)
4	+6.4% +/- 2.0
6	+10.2% +/- 1.9
8	+20.5% +/- 1.9

test-backend-ops -o MUL_MAT: 1288/1288.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: Claude Opus helped me write the sweep script and implement the PR.