Add masked expanded loads and compressed stores API

[gnzlbg]

There is an LLVM patch under review (https://reviews.llvm.org/D59180) that adds portable support for masked compressed loads and masked expand stores via llvm.masked.compressstore and llvm.masked.expandload.

We should add two APIs to vector masks:

impl<T, const N: usize> Simd<[T; N]>  
    where Simd<[T; N]>: Mask 
{
    /// Writes the lanes of `value` for which `self` is `true` to 
    /// adjacent memory locations at `ptr`.
    unsafe fn write_compressed<U>(self, value: Simd<[U; N]>, ptr: *mut U);
    /// Reads adjacent `U` values at `ptr` into the result's vector 
    /// lanes for which `self` is `true`, using the values of `defaults` 
    /// otherwise.
    unsafe fn read_expanded<U>(self, defaults: Simd<[U; N]>, ptr: *const U) -> Simd<[U; N]>;
}