HIP: use upstream rows_per_warp() on RDNA3.5

[liangliangchang]

The RDNA3.5 override returns 32 for J>=64 && J%32==0, with carve-outs holding q4_0/q8_0/q4_K/q5_K at 16. rows_per_warp/16 is ntx, so 32 selects the ntx=2 schedule, and for every type without a carve-out that schedule is several times slower on gfx1151.

This drops the override so RDNA3.5 uses upstream's unconditional 16. The resulting function is byte-identical to ggml-org/llama.cpp master.

Measurements

test-backend-ops perf on gfx1151, m in {1024, 4096}, k=4096. Percent faster with upstream's 16:

type	n=48	n=64	n=96	n=128
iq4_nl	+0.0	+553.1	+441.8	+2.0
iq4_xs	-0.0	+543.1	+484.2	+2.4
mxfp4	-0.0	+539.6	+437.1	+2.6
q5_0	+0.1	+448.6	+418.1	+6.4
iq2_xxs	-0.0	+414.9	+314.5	+1.6
iq3_s	+0.1	+409.5	+291.6	+1.9
q5_1	-0.0	+15.5	+18.2	+8.3
q4_1	-0.0	+3.1	+0.6	+22.7
q6_K	-0.2	+3.5	+0.0	+2.0
q3_K	+0.1	+2.1	+0.1	-0.3
q4_K	+0.0	+0.4	-0.2	+1.3
q5_K	-0.0	+0.8	+0.8	+0.2
q2_K	-0.2	-0.1	-0.1	+0.1
q4_0 (control)	+0.0	+0.0	-0.0	-0.0
q8_0 (control)	+0.2	-0.1	-0.0	-0.0

Six types run 4-7x slower today at n=64 and n=96.

Controls and reproducibility

• n=48 is a control: both variants return 16 there.
• q4_0 and q8_0 are a second control, already 16 at every width.
• All control cells stay within +-0.2%.
• Every effect above 5% reproduces on a second board, same sign and within 2x, at a per-shape noise floor of 0.06-0.74%.

Coverage note

The stock perf list jumps from n=32 straight to n=128, so J=64 and J=96 - the only widths where q4_K/q5_K differ - are not covered by it. These numbers come from an added n in {48, 64, 96, 128} grid; that harness change is not part of this PR.

Q4_K_M models are unaffected end to end (+-0.5% on Qwen3-8B, Qwen3.5-9B, LFM2-8B-A1B), because q4_K/q5_K/q6_K are exactly the types whose deltas are under 3.5%. The large wins land in the iq*, mxfp4, q5_0 and q5_1 families.

Why not a finer-grained rule

One shape prefers 32 - q4_K m=4096 n=96, at -1.1% - against 35 shapes where 16 wins by more than 5%. Keeping a per-type rule is not worth re-creating the hazard.

Correctness

1510/1510 MUL_MAT on both arms.

Assisted-by: Claude Opus 5
Co-authored-by: Cursor cursoragent@cursor.com