No obvious way to xor byte stream with [u8; 4]

I'm trying to get into SIMD by implementing a trivial operation: XOR unmasking of a byte stream as required by the WebSocket specification. The implementation in x86 intrinsics is actually [very straightforward](https://stackoverflow.com/a/17743749/585725), but I have a hard time wrapping my head around expressing it in terms of Faster iterators API.

The part I'm having trouble with is getting an input `[u8; 4]` to cycle within a SIMD vector of `u8`. I have looked at:

1. [`load()`](https://docs.rs/faster/0.4.3/faster/vecs/struct.u8x16.html#method.load) which does accept `&[u8]` as input, but its behavior in case of length mismatch is completely undocumented. It's also not obvious what `offset` parameter does.
1. Casting the input `[u8; 4]` to `u32`, calling `vecs::u32s()` and then downcasting repeatedly to get a SIMD vector of u8, but [Downcast](https://docs.rs/faster/0.4.3/faster/intrin/trait.Downcast.html) seems to do not at all what I want.
1. Getting a SIMD vector of length 4 and arbitrary type inside it, load `[u8; 4]` into it (lengths now match, so it should work) then downcast repeatedly until I get a vector of u8 with arbitrary length. Except there seems to be no way to request a SIMD vector of length 4 and arbitrary type.
1. After over an hour of head-scratching I've noticed that `From` is implemented for `u8x16`, so I could replace Downcast with it in approach 2 and probably get the correct result, except I have no idea how such conversions interact with host endianness.

I actually expected this to be a trivial task. I guess for someone familiar with SIMD it is, but for the likes of me a snippet in examples/ folder that loads `[u8; 4]` into a vector would go a long way. Or perhaps even a convenience function in the API that deals with endianness properly, to make it harder to mess up.