llama: end-to-end tests (#19802)

* tests: add end-to-end tests per model architecture

* fixup for rebase

* fix use-after-free in llama-model-loader.cpp

* fix CI

* fix WebGPU

* fix CI

* disable CI for macOS-latest-cmake-arm64

* use expert_weights_scale only if != 0.0f

* comments

src/llama-arch.h

@@ -4,6 +4,7 @@
 
 #include <string>
 #include <set>
+#include <vector>
 
 //
 // gguf constants (sync with gguf.py)
@@ -608,6 +609,8 @@ struct llm_tensor_info {
     ggml_op op;
 };
 
+std::vector<llm_arch> llm_arch_all();
+
 const char * llm_arch_name(llm_arch arch);
 
 llm_arch llm_arch_from_string(const std::string & name);