ggml-rocket: min_batch floor — keep M<32 (decode) on CPU

llama-bench A/B on boltzmann (gemma-4-E2B-Q8_0, A76-pinned, -t 4) showed the
NPU path HALVED decode: tg128 CPU 7.97 vs Rocket 3.88 t/s, because every M=1
matmul paid the full per-tile BO-alloc/mmap/submit/wait/close round-trip while
decode is bandwidth-bound (the vendor rknpu2 campaign found the NPU only +7%
there even at full offload). Gate supports_op on M>=ROCKET_MIN_BATCH(32),
mirroring ggml-blas, so small batches stay on CPU.

After: pp512 40.32 (was 40.52, +12% vs CPU 35.95), tg128 8.34 (was 3.88, now
>= CPU). Strictly >= CPU on both phases instead of a 2x decode regression.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EWpfhDgYNA21tETDP9ueBE

ggml/src/ggml-rocket/ggml-rocket.cpp

@@ -40,6 +40,7 @@ extern "C" {
 #define ROCKET_TILE_M 16
 #define ROCKET_TILE_N 128
 #define ROCKET_K_MAX  8192
+#define ROCKET_MIN_BATCH 32   // below this, CPU wins (decode is bandwidth-bound)
 
 struct ggml_backend_rocket_context {
     int fd = -1;
@@ -350,7 +351,12 @@ static bool ggml_backend_rocket_device_supports_op(ggml_backend_dev_t dev, const
         case GGML_OP_MUL_MAT:
         {
             const int64_t K = src1->ne[0]; // == src0->ne[0]
+            const int64_t M = op->ne[1];   // batch rows (tokens)
 
+            // Offload only when the NPU can amortise its per-tile submit
+            // overhead: decode (M=1) is bandwidth-bound and strictly faster on
+            // CPU, so keep small batches there (mirrors ggml-blas min_batch).
+            if (M < ROCKET_MIN_BATCH)             return false;
             if (src1->type != GGML_TYPE_F32)      return false;
             if (!ggml_is_contiguous(src0))        return false;
             if (!ggml_is_contiguous(src1))        return false;