HIP: drop the RDNA3.5 q8_0 J=64 -> J=96 widening

The widening existed because J=64 was several times slower for q8_0, but that was
the generic ntx=2 schedule which rows_per_warp() no longer selects for block
quants. With J=64 healthy the widening only pads 64 valid columns into a 96-wide
tile.

Cosmos-Reason2-8B Q8_0, llama-bench -p 512 -r 3 on gfx1151:

  ubatch  64 (J 96 -> 64): 883.1 -> 1118.5 t/s  (+26.7%)
  ubatch  96 (J 96, same): 1151.4 -> 1147.2 t/s (-0.4%)
  ubatch 128 (J 128, same): 1416.1 -> 1398.1 t/s (-1.3%)

MoE keeps its own widening, which is untested here and unchanged.
test-backend-ops MUL_MAT: 1484/1484.

Co-authored-by: Cursor <cursoragent@cursor.com>

Author: liangliangchang

ggml/src/ggml-cuda/mmq.cuh

@@ -1822,20 +1822,21 @@ static bool mmq_rdna35_dual_wg_eligible(const ggml_type type) {
 
 // J values that measure faster on RDNA3.5 than the width the occupancy rule picks.
 //
-// J=64 is 6-7x slower than the neighbouring widths for q8_0; the other types stay within
-// 1.05x of their best there and keep it. For MoE the compacted per-expert grid already supplies
-// the parallelism that column tiling supplies for a dense GEMM, so anything above a 32-wide tile
-// only adds padding - q8_0 excepted, whose narrow tiles are slow.
+// Dense shapes now keep the width the occupancy rule picks: q8_0 used to be widened from
+// J=64 to J=96 because J=64 was several times slower, but that was the generic ntx=2
+// schedule, which rows_per_warp() no longer selects for block quants. Without the widening
+// q8_0 avoids padding 64 valid columns into a 96-wide tile and is 27% faster there.
+//
+// For MoE the compacted per-expert grid already supplies the parallelism that column tiling
+// supplies for a dense GEMM, so anything above a 32-wide tile only adds padding - q8_0
+// excepted, whose narrow tiles are slow.
 static int mmq_rdna35_tuned_J(const ggml_type type, const bool moe, const int J_occupancy) {
     if (moe) {
         if (type == GGML_TYPE_Q8_0) {
             return J_occupancy > 32 ? 96 : J_occupancy;
         }
         return J_occupancy < 32 ? J_occupancy : 32;
     }
-    if (J_occupancy == 64 && type == GGML_TYPE_Q8_0) {
-        return 96;
-    }
     return J_occupancy;
 }
 #endif // GGML_USE_HIP