models : move build_arch_graph() after graph() template specialization

[cpeterso]

Overview

Move build_arch_graph()'s function definitions after the graph<true> and graph<false> template specializations have been explicitly defined.

Additional information

I get the following build errors when compiling llama.cpp as C++23 in my application. C++23 makes std::unique_ptr constexpr, which causes clang to eagerly instantiate std::unique_ptr's body when the explicit specialization of the graph template hasn't been defined yet. This std::unique_ptr change came from C++23 proposal P2273R3:

https://open-std.org/jtc1/sc22/wg21/docs/papers/2021/p2273r3.pdf

In dflash.cpp, eagle3.cpp, and t5.cpp, build_arch_graph() calls std::make_unique, which implicitly instantiates the specializations of the nested graph templates before the explicit specializations are defined. To fix the errors, move build_arch_graph() to the end of each file, after all explicit specializations.

llama.cpp/src/models/dflash.cpp:276:34: error: explicit specialization of 'graph' after
      instantiation
  276 | llama_model_dflash::graph<true>::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
      |                                  ^
/Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/c++/v1/__memory/unique_ptr.h:759:30: note: 
      implicit instantiation first required here
  759 |   return unique_ptr<_Tp>(new _Tp(std::forward<_Args>(__args)...));
      |                              ^

llama.cpp/src/models/eagle3.cpp:137:34: error: explicit specialization of 'graph' after
      instantiation
  137 | llama_model_eagle3::graph<true>::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
      |                                  ^
/Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/c++/v1/__memory/unique_ptr.h:759:30: note: 
      implicit instantiation first required here
  759 |   return unique_ptr<_Tp>(new _Tp(std::forward<_Args>(__args)...));
      |                              ^

llama.cpp/src/models/t5.cpp:122:31: error: explicit specialization of 'graph' after
      instantiation
  122 | llama_model_t5::graph<false>::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
      |                               ^
/Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/c++/v1/__memory/unique_ptr.h:759:30: note: 
      implicit instantiation first required here
  759 |   return unique_ptr<_Tp>(new _Tp(std::forward<_Args>(__args)...));
      |                              ^

Requirements

• I have read and agree with the contributing guidelines: YES
• AI usage disclosure: YES. I used Codex to decipher the C++ error and learn why it wasn't a problem until C++23. I wrote the commit message and this PR myself.