Refactor Idea: wtype per tensor from file instead of global #608

Job	Run time
macOS-latest-cmake	1m 47s
ubuntu-latest-cmake	1m 35s
windows-latest-cmake (noavx, -DGGML_AVX=OFF -DGGML_AVX2=OFF -DGGML_FMA=OFF -DSD_BUILD_SHARED_LIBS...	2m 8s
windows-latest-cmake (avx2, -DGGML_AVX2=ON -DSD_BUILD_SHARED_LIBS=ON)	2m 23s
windows-latest-cmake (avx, -DGGML_AVX2=OFF -DSD_BUILD_SHARED_LIBS=ON)	1m 22s
windows-latest-cmake (avx512, -DGGML_AVX512=ON -DSD_BUILD_SHARED_LIBS=ON)	2m 23s
windows-latest-cmake (cuda12, -DSD_CUBLAS=ON -DSD_BUILD_SHARED_LIBS=ON)	1h 19m 40s
windows-latest-cmake (rocm5.5, -G Ninja -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DS...	27m 30s
windows-latest-cmake (vulkan, -DSD_VULKAN=ON -DSD_BUILD_SHARED_LIBS=ON)	2m 26s
release	0s
	2h 1m 14s

Provide feedback