libclc should remove __CLC_SUBNORMAL_DISABLE and reimplement __clc_fp32_subnormals_supported

[arsenm]

This does not need to be a library build or link time constant, but can be performed with a runtime check which will specialize to a constant based on the ultimate use.

The build should compile with -Xclang -fdenormal-fp-math=dynamic, and then replace __clc_fp32_subnormals_supported with a check like:

  __builtin_isfpclass(__builtin_canonicalizef(0x1p-149f), __FPCLASS_POSZERO)

__clc_flush_denormal_if_not_supported is also approximately equivalent to llvm.canonicalize, but implementing it as
__builtin_fabsf(x) < 0x1p-149f ? copysign(0.0f, x) : x avoids the requirement to quiet a signaling nan so is slightly more foldable.