cuda : improve text-generation and batched decoding performance (#3776)

* cuda : prints wip

* cuda : new cublas gemm branch for multi-batch quantized src0

* cuda : add F32 sgemm branch

* cuda : fine-tune >= VOLTA params + use MMQ only for small batches

* cuda : remove duplicated cuBLAS GEMM code

* cuda : add CUDA_USE_TENSOR_CORES and GGML_CUDA_FORCE_MMQ macros

* build : add compile option to force use of MMQ kernels

llama.h

@@ -178,7 +178,7 @@ extern "C" {
         float rope_freq_scale; // RoPE frequency scaling factor, 0 = from model
 
         // Keep the booleans together to avoid misalignment during copy-by-value.
-        bool mul_mat_q;  // if true, use experimental mul_mat_q kernels
+        bool mul_mat_q;  // if true, use experimental mul_mat_q kernels (DEPRECATED - always true)
         bool f16_kv;     // use fp16 for KV cache, fp32 otherwise
         bool logits_all; // the llama_eval() call computes all logits, not just the last one
         bool embedding;  // embedding mode only