Remove accidental cargo knowledge from `CrateGraph` deduplication

[Veykril]

.., but I would say this is a significant architectural bug.

By design, this layer of rust-analyzer knows nothing about cargo specific concepts. dev-vs-build-vs normal is 100% Cargo concept. rustc knows nothing about these words. As such, any special handling of these concepts should happen in workspace.rs, not here.

The motivation for this design is two-fold:

• practically, Cargo is a build system, there might be others. In the core analysis parts, we shouldn't be prioritizing Cargo over other (current and future) build systems
• theoretically, it is important that the internal objects in rust-analyzer's semantic model reflect the physical reality of rustc. In that physical reality, cargo packages do not exists (and -dev as a concept applies to a package), there are only units of compilation.

Originally posted by @matklad in #15754 (comment)

[alibektas]

@rustbot claim

[alibektas]

So what has been bugging me with this issue is

1. We need to filter what is from Cargo's perspective a DevDependency out of a crate's dependencies so that our new deduplication mechanism works properly.
2. DevDependency is a cargo property this as mentioned in the issue. So it is not a base-db matter it belongs to a less abstract layer.
3. We do not have a multi-phased processing for crates : what we have at hand is immediately turned into a CrateData.

How can I ever make queries like "Get me all the dev-dependencies on this CrateData on this CrateGraph if CrateGraph is the one and only representation we have at our disposal? The only solution I can see to this problem is to have an intermediary phase that acts as a bridge between project-model and base-db::input and is 100% build system ( cargo ) specific.

[Veykril]

(repeating from dms)

Idea would be to have Workspace::to_crate_graph spit out an additional mapping that reveals dev-dependencies in a some fashion (empty for anytihng but cargo workspaces), then CrateGraph::extend can take a call back that looks something like impl Fn(&CrateData, &CrateData) -> Option<CrateData> which is responsible for doing the merge/deduplication so the caller can decide whether to do so. That's where the map then comes in. That should allow lifting the knowledge out of the CrateGraph.

[Veykril]

On a related note, CrateOrigin kind of has a similar issue, as it contains repo: Option<String> which is technically speaking also cargo specific I think (iirc we added that for lsif).