feat: do not convert more flux tensors #506

Job	Run time
macOS-latest-cmake	1m 19s
ubuntu-latest-cmake	1m 18s
windows-latest-cmake (noavx, -DGGML_AVX=OFF -DGGML_AVX2=OFF -DGGML_FMA=OFF -DSD_BUILD_SHARED_LIBS...	1m 23s
windows-latest-cmake (avx2, -DGGML_AVX2=ON -DSD_BUILD_SHARED_LIBS=ON)	1m 44s
windows-latest-cmake (avx, -DGGML_AVX2=OFF -DSD_BUILD_SHARED_LIBS=ON)	1m 16s
windows-latest-cmake (avx512, -DGGML_AVX512=ON -DSD_BUILD_SHARED_LIBS=ON)	1m 20s
windows-latest-cmake (cuda12, -DSD_CUBLAS=ON -DSD_BUILD_SHARED_LIBS=ON)	1h 12m 12s
windows-latest-cmake (rocm5.5, -G Ninja -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DS...	21m 19s
release	25s
	1h 42m 16s

Provide feedback