[FEA]: Add floating point manipulation helpers to scale fp values

In our new improved math implementations, we often need to scale down floating point values below a given boundary to avoid precision issues.

Given that those are nontrivial to read we should provide optimized implementations of effectively `frexp` / `ldexp` to allow easy to use scaling of those fp values.

_Originally posted by @davebayer in https://github.com/NVIDIA/cccl/pull/7632#discussion_r2795152730_