add resolve only pass

[ZacLN]

Adds a new option for semantic_pass whereby only 'target' expressions get the full treatment (i.e. build new scopes, add bindings etc) for the rest we only resolve references. The logic isn't fully worked out but the idea is that we'd only run everything on the toplevel scope and those parts of the tree (delayed scope - e.g. functions) that are new - for the rest we only need to ensure that any new bindings that have been added to the toplevel scope are resolved.

[ZacLN]

Some timings:

using CSTParser, StaticLint, SymbolServer, BenchmarkTools
server = StaticLint.FileServer()
ssi = SymbolServerInstance("/your/depot/path", "/your/cache/path")
_, server.symbolserver = SymbolServer.getstore(ssi, "~/.julia/environments/v1.7/")
server.symbol_extends  = SymbolServer.collect_extended_methods(server.symbolserver)


function setup(rootp, fp, server)
    empty!(server.files)
    root = StaticLint.loadfile(server, rootp)
    StaticLint.semantic_pass(root)
    f = StaticLint.getfile(server, fp)
    root, f
end

begin
    root, f = @btime setup(Base.find_source_file("Base.jl"), Base.find_source_file("abstractarray.jl"), server);
    @btime (StaticLint.clear_meta.(CSTParser.EXPR[f.cst[32]]);StaticLint.semantic_pass(root))
    @btime (StaticLint.clear_meta.(CSTParser.EXPR[f.cst[32]]);StaticLint.semantic_pass(root, CSTParser.EXPR[f.cst[32]]))
    @btime StaticLint.semantic_pass(f)
end
# 723.366 ms (3663143 allocations: 229.04 MiB)
#  171.288 ms (253585 allocations: 15.83 MiB)
#  86.797 ms (81905 allocations: 4.32 MiB)
#  4.115 ms (5311 allocations: 395.45 KiB)

So we get about a 2x speedup on handling edits. This can probably be sped up further if we feed in information on what has actually changed at the toplevel (i.e. with this approach we're still trying to resolve anything that hasn't got a reference - we could limit that by only trying to resolve symbols matching new names that have been added).

[ZacLN]

Subject to julia-vscode/CSTParser.jl#224 being merged, and a new verison tagged, this should be good to go.

[davidanthoff]

Don't understand enough to really review this, but superficially it all looks great :) So, I'd say we merge.

Does this depend on a new CSTParser version? Shouldn't some test fail in that case?