Document that it is okay to temporarily have invalid data behind an `&mut T`

[RalfJung]

As part of this discussion, it seems we have general consensus that "it is acceptable to write bytes through a pointer derived from &mut T even if those bytes make the T invalid so long as the T is made valid again before the lifetime that would be given to the pointer if it were a reference expires and the parent &mut T is usable again."

Specifically, this is fine:

fn main() {
    let mut b = false;
    let ptr = &mut b as *mut bool as *mut u8;
    unsafe {
        *ptr = 2;
        *ptr -= 1;
    }
    assert!(b);
}

We should probably document this as UCG consensus, and maybe even find a way to get t-lang to sign off on this, though it seems unclear how to best do that.

This is helpful because it means the code in rust-lang/rfcs#2195 is fine without having to weaken the notion of validity of those enums specifically.

[Darksonn]

This seems like it's just a weaker version of #346.

[RalfJung]

I would call it a weaker version of #84. But yeah it overlaps with both of those discussions - without finally resolving either of them.

[RalfJung]

Closing in favor of #346