[XLA:CPU] Allow users to enable fast-math without enabling nonan and noinf

This CL adds two flags, xla_cpu_fast_math_allow_{nans|infs}, that can be set in
conjunction with xla_cpu_enable_fast_math to enable all fast-math flags except
NoNaN/NoInf.

We don't want these flags to live long-term; once we're happy with the
performance we'll remove them by permanently setting them to true.

As part of this change I added a Cpu substring to some functions to make it
obvious that they're only used by the CPU backend.

PiperOrigin-RevId: 238663934