Tidy up "patch-up" logic.

In the last PR I added the ability to "patch-up later" instruction
indices that are as-yet unknown. I stashed the info needed for this in a
new field in the YKIRWriter.

However, since that info is "per-function", there's no need to keep it
alive that long. In fact the lifetime is the same as the `VLMap`.

This change groups the VLMap and the patch-up info together into a
`FuncLowerCtxt`. A fresh one of these is created for each LLVM function
we lower, and it is then passed down through the serialiser using no
object state from `YkIRWriter`. When we are done lowering a function, we
allow the `FuncLowerCtxt` to fall out of scope, thus better capturing
the lifetime of this information and preventing accidental reuse.

No functional change.

Author: vext01