Compute padding outside of codegen

Right now, we compute a struct's padding fields while we are codegen'ing it.

Instead, we should have a dedicated phase for computing implicit padding fields that transforms from structs-with-only-explicit-fields to structs-with-explicit-fields-and-implicit-padding-fields.

Benefits this will give us:

• The ability to precisely derive various traits for structs with large alignment (right now we are conservatively pessimistic). See for example https://github.com/rust-lang-nursery/rust-bindgen/blob/b13f0766e28cf75dd04f358bf64d87ea96b111c3/src/ir/analysis/derive_debug.rs#L160-L170

• Separation of concerns, simpler code

This new phase should probably happen either right before or right after we compute allocation units for bitfields. I'm not 100% sure which, and determining it requires further investigation/experimentation.

Start by tracing where struct padding is computed during code generation and where bitfield allocation units are computed. Read src/ir/analysis/derive_debug.rs, especially the referenced trait-derivation logic, and investigate whether the new padding phase belongs before or after bitfield allocation. Done means structs are transformed to include implicit padding before later analyses can derive precise traits.