CUDA: adding switch points per HW and quant type to tune the mvq->MMQ decode crossover

ggml/src/ggml-cuda/mmvq.cu

@@ -281,6 +281,42 @@ bool ggml_cuda_should_use_mmvq(enum ggml_type type, int cc, int64_t ne11) {
     if (!ggml_is_quantized(type)) {
         return false;
     }
+    // k-quants cost more to decode and mvq redoes that per column, so MMQ wins sooner.
+    // Only list types MMQ supports, others would fall back to cuBLAS.
+    if (GGML_CUDA_CC_IS_NVIDIA(cc) && cc == GGML_CUDA_CC_ADA_LOVELACE) {
+        switch (type) { // tuned on RTX 4090
+            case GGML_TYPE_Q2_K:
+                return ne11 <= 4;
+            case GGML_TYPE_Q3_K:
+                return ne11 <= 6;
+            case GGML_TYPE_Q4_K:
+            case GGML_TYPE_Q5_K:
+                return ne11 <= 7;
+            default:
+                return ne11 <= MMVQ_MAX_BATCH_SIZE;
+        }
+    }
+    if (GGML_CUDA_CC_IS_NVIDIA(cc) && cc == GGML_CUDA_CC_BLACKWELL) {
+        switch (type) { // tuned on RTX 5090
+            case GGML_TYPE_Q2_K:
+            case GGML_TYPE_Q3_K:
+            case GGML_TYPE_Q4_K:
+            case GGML_TYPE_Q5_K:
+                return ne11 <= 5;
+            case GGML_TYPE_Q6_K:
+                return ne11 <= 7;
+            default:
+                return ne11 <= MMVQ_MAX_BATCH_SIZE;
+        }
+    }
+    if (GGML_CUDA_CC_IS_NVIDIA(cc) && cc == GGML_CUDA_CC_DGX_SPARK) {
+        switch (type) { // tuned on DGX Spark GB10
+            case GGML_TYPE_Q2_K:
+                return ne11 <= 6;
+            default:
+                return ne11 <= MMVQ_MAX_BATCH_SIZE;
+        }
+    }
     if (GGML_CUDA_CC_IS_CDNA(cc)) {
         if (GGML_CUDA_CC_IS_CDNA1(cc)) {
             switch (type) {