CL: Math fixes

Should fix errors reported by the bruteforce test.

You'll need https://gitlab.freedesktop.org/bbrezillon/llvm-project/-/tree/10.x-libclc-windows-dxil to test.

v2:

• Followed @daniels suggestion to use a __builtin and let vtn tweak the value of __builtin_hw_has_fma32 based on the ->lower_fma flag. This way we should always end up with an optimize version of the sincos helpers
• Use clc implementation as a fallback when the NIR implementation is not precise enough (decision based on the ->lower_xxx flags)
• Fix nextafter() for platforms that don't support denorms

/cc @jenatali @daniels