Add option to specify compiler flags on the CPU (like -ffast-math)

[martinloretzzz]

The -ffast-math compiler flag isn't set on the CPU device, it would be cool if one can specify custom compile flags for the CPU device in the build.toml.

The current compile command is the following:

Build Command

`/nix/store/m6s7rszf3mbmgnba8l3bmzn22f4s39c6-gcc-wrapper-13.4.0/bin/g++ -DCPU_KERNEL -DPy_LIMITED_API=3 -DTORCH_EXTENSION_NAME=_slstm_kernels_cpu_9603e67_dirty -DUSE_C10D_GLOO -DUSE_DISTRIBUTED -DUSE_RPC -DUSE_TENSORPIPE -D_slstm_kernels_cpu_9603e67_dirty_EXPORTS -I/nix/var/nix/builds/nix-976535-1632023732/source/slstm_cpu/external/eigen -I/nix/var/nix/builds/nix-976535-1632023732/source/csrc -isystem /nix/store/3lll9y925zz9393sa59h653xik66srjb-python3-3.13.9/include/python3.13 -isystem /nix/store/dmic0d3gwcdakzsq8nfnqyp7fyb3kskh-python3.13-torch-2.10.0/lib/python3.13/site-packages/torch/include -isystem /nix/store/dmic0d3gwcdakzsq8nfnqyp7fyb3kskh-python3.13-torch-2.10.0/lib/python3.13/site-packages/torch/include/torch/csrc/api/include -O3 -DNDEBUG -fPIC -MD -MT CMakeFiles/_slstm_kernels_cpu_9603e67_dirty.dir/slstm_cpu/slstm_cpu.cpp.o -MF CMakeFiles/_slstm_kernels_cpu_9603e67_dirty.dir/slstm_cpu/slstm_cpu.cpp.o.d -o CMakeFiles/_slstm_kernels_cpu_9603e67_dirty.dir/slstm_cpu/slstm_cpu.cpp.o -c /nix/var/nix/builds/nix-976535-1632023732/source/slstm_cpu/slstm_cpu.cpp`

[martinloretzzz]

Fixed by #455, but it might also be a good to enable the -ffast-math by default