vulkan: fix group_norm (llama/10496)

Fix bad calculation of the end of the range. Add a backend test that
covers the bad case (taken from stable diffusion).

Fixes leejet/stable-diffusion.cpp#439.

Author: jeffbolznv

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -7157,7 +7157,7 @@ static void ggml_vk_check_results_0(ggml_tensor * tensor) {
         const int32_t max_period = tensor->op_params[1];
         tensor_clone = ggml_timestep_embedding(ggml_ctx, src0_clone, dim, max_period);
     } else if (tensor->op == GGML_OP_POOL_2D) {
-        enum ggml_op_pool op = static_cast<ggml_op_pool>(dst->op_params[0]);
+        enum ggml_op_pool op = static_cast<ggml_op_pool>(tensor->op_params[0]);
         const int32_t k0 = tensor->op_params[1];
         const int32_t k1 = tensor->op_params[2];
         const int32_t s0 = tensor->op_params[3];

ggml/src/ggml-vulkan/vulkan-shaders/group_norm.comp

@@ -19,7 +19,7 @@ void main() {
 
     const uint tid = gl_LocalInvocationID.x;
     const uint start = gl_WorkGroupID.x * group_size + tid;
-    const uint end = start + group_size;
+    const uint end = (gl_WorkGroupID.x + 1) * group_size;
 
     tmp[tid] = 0.0f;