save-analysis: use a decoupled representation for dumped data

[aochagavia]

Currently, the Dump trait exposes internal types of the compiler, which is undesirable. It also forces the trait's implementors to write lowering code (as can be seen in the current implementation of json_dumper).

Proposed solution

1. Create a new module in save-analysis called external_data where the new *Data structs will live (they can be moved from json_dumper, along with the lowering code).
2. Modify the Dump trait to operate on save_analysis::external_data::* instead of on save_analysis::data::*.

Some open problems

When getting a DefId from a NodeId, what should we do when tcx.map.opt_local_def_id(node_id) returns None? What does that mean? If we call the function with a NodeId provided by the compiler, should it always succeed? The current code doesn't make that assumption (see data.rs)

[aochagavia]

cc @nrc

Do you know the answer to my questions in the open problems part?

[nrc]

@aochagavia if the def id is from the local crate, then it should never return None. I think we can assume that and just unwrap. We do need to make sure we only call with local ids though, it is possible we could get a NodeId from an extern crate (via metadata or something?).