Clarify cross-crate dependency story

[nikomatsakis]

We're currently doing some wrong-ish things with respect to dep-graph tracking across crate boundaries. I think what I want to do is this:

• Create a DepNode::MetaData(DefId) variant that plays a role like Hir
  • it basically represents the metadata for some item
• Whenever we load from metadata for item X, we'll generate a read on MetaData(X)
• In the metadata, we'll include a hash for each item, either:
  • hashing the saved contents of the item
  • or hashing all the base input nodes that are deps of the things serialized into
  • basically some hash we can use to tell if the metadata for an item changed in some way
• When we save the dep-graph, we'll save the hash out of the crate metadata
• When we load, we'll reload the hash out of the current crate metadata and compare

For inlined items, I hope we will transition to MIR and not need to inline sooner or later, but in the meantime, I think we can basically translate reads to Hir(X) where X is an inlined def-id to MetaData(Y) where Y is the original def-id -- we may also just be able to ignore accesses to inlined def-ids, since we'll presumably add an appropriate edge from some other source.

[michaelwoerister]

I think this is implemented (mostly via #32016) except for correctly handling access to inlined HIR.
As far as I can see, there is no special handling of inlined HIR anywhere, which means that the following will happen:

1. When an item gets inlined during trans, we will add a new HIR-depnode for the inline-copy of the item.
2. This inline-copy has a local DefId, thus its hash will be stored with any other HIR dep-nodes in the dep-graph.
3. When loading the dep-graph during the next round, the corresponding AST/HIR is not present, as it has not been inlined yet.
4. As a consequence of (3), everything dependent on the inlined item will be invalidated.

[nikomatsakis]

@michaelwoerister

I think this is implemented (mostly via #32016) except for correctly handling access to inlined HIR.

I agree -- however, it seems like inlined HIR is going away. That is, once we transition to MIR trans, this point will no longer be an issue, right?