Add tests for signaling NaNs, and fix function behavior for handling
signaling NaN input.

Fixes https://github.com/llvm/llvm-project/issues/124812

For targets that have single precision FPU but not double precision FPU
such as Cortex M4, only using float-float in the intermediate
computations might reduce the code size compared to using double. In
this case, when the exact pass is skipped, the float-only option for
atan2f implemented in this PR reduces the code size of this function by
~1 KB compared to the double precision version.

This is a part of #97655.

Reverts llvm/llvm-project#98075

bots are broken

This is a part of #97655.

We compute atan2f(y, x) in 2 stages:
- Fast step: perform computations in double precision , with relative
errors < 2^-50
- Accurate step: if the result from the Fast step fails Ziv's rounding
test, then we perform computations in double-double precision, with
relative errors < 2^-100.

On Ryzen 5900X, worst-case latency is ~ 200 clocks, compared to average
latency ~ 60 clocks, and average reciprocal throughput ~ 20 clocks.