[XLA:GPU] Force-enable two math transformations that previously were only on with fastmath.

  - pow(x, -.5) => 1/sqrt(x)
  - tanh(x) => fast llvm-ir implementation, instead of __nv_tanh from
    libdevice.

Both of these should be safe.  The main thing we're concerned about when
fastmath is disabled is getting proper NaN/Inf behavior, and these
changes don't affect that.

PiperOrigin-RevId: 208511562