Add "common" shuffles

[calebzulawski]

The shuffle API in #62 can handle nearly any hardware shuffle, but it's a bit clumsy to use (and the API requires full const generics).
A few common cases of shuffles should be provided as a simpler API:

• reverse
• rotate
• shift (like rotate but insert 0s)
• align (or maybe rotate2?) see Are the alignr/alignl simd functions planned? #78
• interleave/deinterleave

[boomshroom]

If the shift amount is a constant, then it should be possible to implement these as const functions that output the corresponding array to passe into shuffle's const parameter.

As an example, it's possible to implement alignr (x86 doesn't have an alignl) with

const fn alignr<const LANES: usize>(shift: usize) -> [u32; LANES] {
    let mut indices = [0; LANES];
    let mut block = 0;
    while block < LANES {
        let mut idx = 0;
        while idx < 16 && idx < LANES {
            // x86 chunks its vectors into 16 byte chunks for alignr
            let offset = if shift + idx >= 16 {
                LANES + (shift + idx) % 16
            } else {
                shift + idx
            };
            indices[idx + block] = (offset + block) as u32;
            idx += 1;
        }
        block += 16;
    }
    indices
}

After some testing, LLVM does appear to recognize the array that results as an alignr and simplifies it down.

[calebzulawski]

I wonder if there's any point adding shift and align? They can both be implemented manually, or they can be implemented via rotate and select, which I would think the compiler could optimize. They're not the most obvious functions, so it might actually be clearer manually implementing them than having to reference the docs.