Use core::intrinsics::powi{f32,f64} for powi instead of num-traits

[mikwielgus]

This undoes 8eecf61 code changes, except for the powi.rs test, which is retained.

To provide powi in a no-std environment, instead of pulling it from the num-traits crate and defining a custom intrinsic for it, we just define powi methods on f32 and f64 directly. In these methods, we just emit core::intrinsics::powif32 and core::intrinsics::powif64, respectively.

We can do this thanks to the #[rustc_allow_incoherent_impl] internal rustc attribute, which allows us to add inherent impl blocks on types defined in other crates, which would otherwise violate coherence rules.

(This trait is similarly used inside the Rust compiler to have core module provide some alternative implementations of std's methods in a no-std environment.)

In a normal Rust library, the use of such an internal rustc attribute would have been unacceptable, but since Rust-GPU is a compiler, it seems reasonable.

Unlike f32 and f64, there is no need to implement powi for f16 and f128 because these already have powi methods correctly implemented in core (see https://doc.rust-lang.org/core/primitive.f16.html#method.powi and https://doc.rust-lang.org/core/primitive.f128.html#method.powi). And f32 and f64 actually will have these too, once Rust issue #137578 (rust-lang/rust#137578) is resolved, rendering this commit obsolete. But noone knows how many years it will take them to get that going, so let's just fix the problem on Rust-GPU's end.

There are other mathematical methods, e.g. powf, sqrt, and many others, that Rust-GPU still needs num-traits for. I will be moving their implementations to Rust-GPU and removing num-traits altogether in subsequent commits, after and if this PR is accepted -- this is just the first step to see if this solution is acceptable at all.

I ran cargo test, cargo compiletest, cargo difftest. Tests pass. powi.rs test's output changed, but only marginally -- I have blessed it.

Reference: #520

[LegNeato]

Neat, I did not know about rustc_allow_incoherent_impl!

[Firestar99]

Thanks for the patch, looks good from my side!