Fix struct_extensions projections on mapped and container types

[cpubot]

struct_extensions currently uses field.target_resolved() to determine member projection types. This is incorrect for cases where a struct member is annotated with a container type, as target_resolved will return the fully inferred container type, rather than the underlying type.

For example

struct A {
    #[wincode(with = "containers::Pod<_>")]
    signature: [u8; u32]
}

target_resolved() returns containers::Pod<[u8; 32]>, whereas we actually want the [u8; 32] directly.

Note that target_resolved is still the correct type to use for read_<field>, as read should use the type's read implementation directly (and read_<field> implementation is unchanged). But for write_<field> or uninit_<field>_mut, we need to use the destination type (Dst) since they circumvent the SchemaRead::read impl.

The fix here is simple -- just use <field.ty as SchemaRead>::Dst rather than field.target_resolved(). This works for mapped types, container types, and plain built-in types.

Example mapped type

struct A {
    signature: [u8; u32]
}

#[derive(SchemaWrite, SchemaRead)]
#[wincode(from = "A", struct_extensions)]
struct AMapped {
    signature: containers::Pod<[u8; u32]>,
}

Added tests that will fail to compile on current master.

[kskalski]

could use int_uninit_mut() here

[cpubot]

e38f188

[kskalski]

since there is a non-trivial SchemaRead lifetime handling, could you add a test that works on struct with lifetimes, say

struct TestRef<'a> {
    a: &'a [u8],
    b: Option<'a u8>
}

it should be possible to initialize it with a (locally scoped) slice, right?

[cpubot]

Sure, 5ed3c6b

[kskalski]

Interestingly, tests show UB now?

[cpubot]

Interestingly, tests show UB now?

Yeah, looks like miri didn't like ptr::read before moving self into mem::forget. Fixed with ManuallyDrop 4914cd1