x/tools/gopls: diff metadata before invalidating on changes to go.mod

[stamblerre]

Right now, gopls invalidates all workspace metadata if there is a change to the user's go.mod file. This is unnecessary if the metadata remains valid after the change. @heschik proposes running go list -m all as a gate during the snapshot cloning to determine if we should invalidate all metadata or keep it in the cache. Either way, we will either invalidate all metadata for the workspace or none.

[findleyr]

CC @adonovan

I think this is less urgent now that we only invalidate on saves to go.mod, go.work, etc., but also we're pretty close to being able to do this after recent refactorings to invalidation and loading. It would be easy to avoid applying metadata changes if no diff was detected in cache.buildMetadata.

The problem is getting the package invalidation correct to avoid layering problems: when we invalidate metadata in snapshot.clone, we also invalidate computed packages. With the current caching model, these packages get evicted ~immediately, so even if metadata has not changed following the load they need to be recomputed. If we don't invalidate immediately, then we have packages in the snapshot whose metadata does not match the stored metadata. As we have seen, maintaining such invariants is crucial.

I think the cleanest way to achieve this may be to not do metadata invalidation in snapshot.clone, but just schedule a reload (c.f. https://go.dev/cl/417576). Then, create a new snapshot on load. This has the advantage of making metadata in the snapshot fully immutable. Along with #42814, we can move toward a model where the snapshot is entirely immutable, save for lazily computed promises.

Moving this to the gopls/later milestone.