Simplify and co-locate more completion abstractions in server

[zachallaun]

This PR attempts to make completion code easier to understand and work on.

Previously, code that explicitly referred to completions lived in common, server, and remote_control; after this PR, it all lives in server and remote_control. Because server can refer to protocol types, I was also able to dramatically simplify the types/specs used in the builder.

Additionally, two behaviours have been removed in favor of referring to their implementations directly: Lexical.Ast.Environment (which was only implemented by Lexical.Ast.Env), and Lexical.Completion.Builder (which was only implemented by Lexical.Server.CodeIntelligence.Completion.Builder). Removing some of this indirection also allowed Dialyzer to catch a few bugs in the types that lived in Lexical.Ast.Environment (now Lexical.Ast.Env).

[Moosieus]

Found this from a conversation a few months ago:

I remember why builder is a behaviour, plug-ins. If it wasn't, we'd have to marry releases of lexical-plugin to releases of lexical, and move builder code to lexical-plugin.

Not sure if the same applies to Lexical.Ast.Env though.

[zachallaun]

@Moosieus Hmm, I'm not sure I'm following that. lexical_plugin only has a dependency on lexical_shared, and lexical_shared is the "root" so to speak: it has no dependencies on any other part of Lexical. That should mean it's safe to move around anything in apps/*, as nothing in projects/* depends on them.

[scohen]

It's the other way. The intention behind these behaviours was that we could then move things to lexical_shared, publish the behaviour there and have lexical_shared provide an implementation, and then lexical would provide the real implementation.

[zachallaun]

It's the other way. The intention behind these behaviours was that we could then move things to lexical_shared, publish the behaviour there and have lexical_shared provide an implementation, and then lexical would provide the real implementation.

While I think that's reasonable, I think that this simplification is worth doing for now for a few reasons:

• We haven't moved it to lexical_shared, and to my knowledge it isn't anyone's priority
• There are questions about the utility/design of plugins anyways
• This simplifies things for current contributors at the possible expense of a slightly harder refactoring if we want to turn this into a behaviour later

[scohen]

Yes, i'm generally on board with this simplification, just giving you the context.