Linux CUDA build lacks the fast-math flags the Windows build gets

Found during the fastgs performance review (#1757).

src/training/rasterization/fastgs/CMakeLists.txt adds --use_fast_math only on the MSVC branch. On Linux the kernels compile with full IEEE divisions and transcendental calls. Consequences:

- Windows and Linux builds train with different numerics today, so results and gradient-check behavior are not identical across platforms.
- Linux pays roughly 10 instructions per division; #1757 removed the worst offenders by hand, but many remain.

Decision needed: align the platforms one way or the other. Enabling fast-math on Linux is a one-line change but alters gradients (approximate exp/div, flush-to-zero) and would need the FD gradcheck suite re-validated; removing it from Windows keeps exact math everywhere at some Windows speed cost. Either way, the two builds should match.

[]