[Blocked] Use const fn where easily applicable.

[phip1611]

Update 2023-12-18

This is still blocked by having these features not in stable.

#![feature(const_float_bits_conv)]
#![feature(const_fn_floating_point_arithmetic)]
#![feature(const_float_classify)]

---

This commit adds "const fn" where possible without major refactorings. All functions that rely on special LLVM runtime intrinsic (such as SSE) or core::ptr::read_volatile are untouched by this.

Currently, the CI does not pass. I'd like to know if you think this PR is beneficial at all. If so, we/I can work further on it. Currently, many functions can't be ported to be a "const fn" easily because many of them rely on core::ptr::read_volatile during runtime.

This will be breaking because it requires a recent rust nightly. Let me know what you think!

[Amanieu]

I am generally in favor of the PR!

The volatile reads are a bit of a hack. I believe they were used for 2 purposes:

• To set the appropriate FP status flags (overflow, underflow, etc) in the FP status register. We really shouldn't bother with this since we don't guarantee any particular value for flags.
• To work around FP precision issues on 32-bit x86 which uses 80-bit float precision internally. This is unfortunately an unavoidable issue on x86 since it has pretty broken FP support.