ggml-webgpu: move from parameter buffer pool to single buffer with offsets (#21278)

* Work towards removing bitcast

* Move rest of existing types over

* Add timeout back to wait and remove synchronous set_tensor/memset_tensor

* move to unpackf16 for wider compatibility

* cleanup

* Remove deadlock condition in free_bufs

* Start work on removing parameter buffer pools

* Simplify and optimize further

* simplify profile futures

* Fix stride

* Try using a single command buffer per batch

* formatting

ggml/src/ggml-webgpu/ggml-webgpu-shader-lib.hpp

@@ -437,12 +437,18 @@ inline uint32_t ggml_webgpu_flash_attn_pick_vec_ne(const ggml_webgpu_flash_attn_
 
     // Head-dim specializations used by the tuned vec f16 path.
     switch (key.head_dim_qk) {
-        case 64: return 2u;
+        case 64:
-        case 96: return 4u;
+            return 2u;
-        case 128: return 1u;
+        case 96:
-        case 192: return 2u;
+            return 4u;
-        case 576: return 2u;
+        case 128:
-        default: return 1u;
+            return 1u;
+        case 192:
+            return 2u;
+        case 576:
+            return 2u;
+        default:
+            return 1u;
     }
 }
 
@@ -1857,8 +1863,7 @@ class ggml_webgpu_shader_lib {
         defines.push_back(std::string("SG_MAT_K=") + std::to_string(context.sg_mat_k));
 
         uint32_t q_tile  = context.sg_mat_m;
-        uint32_t kv_tile =
+        uint32_t kv_tile = std::min(ggml_webgpu_flash_attn_max_kv_tile(context),
-            std::min(ggml_webgpu_flash_attn_max_kv_tile(context),
                                     context.sg_mat_n * GGML_WEBGPU_FLASH_ATTN_PREFERRED_KV_SG_TILES);
         if (context.key.use_vec) {
             q_tile  = 1;