#322 redux

Arguments for this change:

• Integer elements impose a very specific view of masks (AVX512, SVE, RVV, etc use bitmasks). Mask<i32, 4> over Simd<f32, 4> only maps to one type of implementation. Mask<f32, 4> just means "a mask over f32s"
• Some API surface is reduced: T::Mask disappears from everywhere except Mask::to_simd/Mask::from_simd. MaskElement is removed from the public API.

This comment from @thomcc summarizes the arguments against this change in the previous PR: #322 (comment)

• I think the first comment regarding API friction is no longer relevant, as this PR includes From implementations for all masks. Additionally, Select was changed in Remove poor-performing bitmasks, add Select trait, and enable select on integer bitmasks #482 to allow any mask to select any vector as long as it has the same number of elements. Some code will require slightly more calls into()/cast() in some situations but otherwise there should be minimal overhead.
• Regarding the second comment about complicated types and projections, this PR doesn't really make this worse anywhere, and probably makes it better in a lot of places, e.g. you can make a function generic over T: SimdElement and that will also be your mask type.