Add PME backstop to detect padding in IntoBytes types

[joshlf]

As demonstrated in #1764, #1763, #1758, and #1757, reasoning about the presence or absence of padding is subtle.

When a type is Sized and concrete, we can easily emit a padding check using size_of to confirm that the size of the type is equal to the sum of the sizes of its fields (although note that this is still unsound in the presence of niche-filling). However, on unsized or generic types, our job is much harder, and #[derive(IntoBytes)] employs subtle logic to determine whether Rust could in principle emit padding bytes in a particular type.

When the type is generic, the only reason that we can't emit a padding check is that it requires the unstable generic_const_exprs feature. However, we could still use a post-monomorphization error (PME). In particular, I propose the following:

• Add a #[doc(hidden)] associated const to IntoBytes which evaluates to () if a type has no padding or panics during evaluation if it does have padding
• Evaluate this const in all IntoBytes methods, and maybe in other places that consume T: IntoBytes

Note that this is a backstop, but it must not be taken as a complete proof that a type is padding-free. In particular, user code is permitted to use T: IntoBytes as a precondition to justify the soundness of unsafe code - we cannot guarantee that such user code will call our methods. Thus, we cannot guarantee that such user code would trigger a PME.