Fix acos op decomposition for non complex numbers to be more accurate

[GleasonK]

What happened?

An initial fix was made for accuracy but was reverted due to a bug in the pattern generator:
#2449

This ticket tracks fixing that work and submitting a more accurate decomposition for acos op.

Current:
// Expand acos for non-complex arguments to MHLO dialect as follows:
//   acos(x) = 2 * atan2(sqrt(1 - x^2), (1 + x))  if x != -1

Using 1-x^2 will be inaccurate when abs(x) is close to 1. The corresponding cancellation errors are avoided when using (1-x)*(1+x).

Steps to reproduce your issue

No response

Version information

No response