x/tools/gopls: optimize calculation of the workspace symbol index

[findleyr]

Right now, the workspace symbol index is populated on the first workspace symbol request, theoretically to optimize for clients that don't use workspace symbols at all.

I recalled this taking ~5s for a large repo like Kubernetes, after which subsequent workspace symbol requests take <100ms. Seemed OK. However, testing this now it often takes 20s+ for the initial population with Kubernetes (and is surprisingly variable). We should either improve this initial performance, or eagerly populate the symbol table.

The symbol handle (which generates symbols for a file) currently bypasses the cache for the parsed file, i.e. re-parses the file, to avoid pinning the parsed file in the cache. We can probably speed up the initial symbol load significantly by supporting cache-peeking, so that we can leverage existing parsed files in memory, if they exist, without pinning.

[findleyr]

In https://go-review.googlesource.com/c/tools/+/403679, I update the symbolizer to

1. Use cached parsed files if they exist.
2. Otherwise parse as quickly as possible (skipping comments and object resolution)

With these two changes, in my experiments the initial symbol load is ~3x faster. (from ~16s for k8s to ~5s, and from 2.4s for x/tools to 700ms.

We could do more and actually pre-populate symbols, but this seems good enough for now.

If we wanted to further improve this, we could populate symbols eagerly when we type-check a package -- we have access to the full AST before we filter it.

[findleyr]

This was fixed by go.dev/cl/52602 (the referenced issue in the Fixes line should have been golang/go#NNNN).