Support derive(KnownLayout) on DSTs

[jswrenn]

Makes progress towards #29. Supersedes #541.

For types marked #[repr(C)], this derive requires that the trailing field is KnownLayout:

// For illustrative purposes; `align` and `packed`
// can't actually be combined like this.
#[repr(C, align(2), packed(4))]
struct ReprCStruct<T, U, V> {
    foo: Foo<T>,
    bar: Bar<U>,
    baz: Baz<V>,
}

impl<T, U, V> KnownLayout for ReprCStruct<T, U, V> {
    const LAYOUT: DstLayout = {
        use ::zerocopy::macro_utils::core_reexport::num::NonZeroUsize;
        use ::zerocopy::{DstLayout, KnownLayout};

        let repr_align = NonZeroUsize::new(2); // or `None`, w/o `align`
        let repr_packed = NonZeroUsize::new(4); // or `None`, w/o `packed`

        DstLayout::new(repr_align)
            .extend(DstLayout::for_type::<Foo<T>>(), repr_packed)
            .extend(DstLayout::for_type::<Bar<U>>(), repr_packed)
            .extend(<Baz<V> as KnownLayout>::LAYOUT, repr_packed)
            .pad_to_align()
    };
}

For non-repr(C) layouts, it's only required that Self: Sized.

[joshlf]

Need to add tests, notably:

• In src/lib.rs, the impl tests for types like Unalign should test with both sized and unsized types
• UI failure tests for the following cases:
  • Type is repr(C), trailing field is a concrete type (not a type parameter), and trailing field doesn't implement KnownLayout
  • Type is not repr(C), trailing field is a concrete type (not a type parameter), and trailing field is unsized
• Non-UI tests for all combinations of the following axes:
  • repr(C)/not repr(C)
  • Trailing field is concrete/generic
  • Trailing field does/does not implement KnownLayout
  • Trailing field is sized/unsized

[joshlf]

For success cases, can we also test that we get the expected value for KnownLayout::LAYOUT?

It's probably also worth adding a test along these lines:

#[repr(C)]
struct Foo<A, B, C>(A, B, C);

macro_rules! test_foo {
    ($a:ty, $b:ty, $c:ty => $layout:expr) => {
        assert_eq!(<Foo<$a, $b, $c> as KnownLayout>::LAYOUT, $layout);
    };
}

// TODO: A bunch of different combinations of types with different sizes and alignments

[joshlf]

Does this need a repr?

[jswrenn]

Discussed offline: It could use a repr(transparent), but it would be sufficiently interesting if this test started failing on a future Rust version, that it'd be nice to get the heads up.