failed to reduce pow() with sqrt() with -ffast-math

```c
#include

double foo(double x) {
return pow(sqrt(x), 4.0);
}
```

This can be reduced to x*x:
https://godbolt.org/z/a78oraYoP

It looks like the special case of pow(N, 2.0) is handled, but all higher exponents are missed. If the pow() is converted to powi() in IR, then there could be another transform to divide the exponent in half?

It also doesn't reduce when the sqrt() follows the pow(), so there's probably a missing pow/sqrt canonicalization:
https://godbolt.org/z/4e1Phv8Y7