perf(lsp): cache lint and doc diagnostics per document module (#34875)
Right now, the LSP regenerates lint diagnostics for every open file each
time diagnostics are refreshed:
https://github.com/denoland/deno/blob/ed3086e4b129843c13a009112cf21dfd05745905/cli/lsp/diagnostics.rs#L451
This means that after enough files are open (say 50), there is a
noticeable delay when changing only a single file, since every open file
is re-linted on each refresh.
Lint diagnostics only depend on a file's own contents — not on other
files — so they can be computed once and reused until the file changes.
This PR caches lint (and `deno doc --lint`) diagnostics on the
`DocumentModule` instance via a `OnceCell`.
### Why caching on the module instance is correct
A new `DocumentModule` is created whenever:
- the document content changes (the module map is keyed by the
`OpenDocument` pointer, so a new version produces a new module), or
- the lint configuration changes (every `refresh_linter_resolver()` call
site is immediately followed by `refresh_documents_config()`, which
resets the module maps and drops all cached modules).
So the module instance is a sound cache key for content-only
diagnostics, and the cache is invalidated automatically when either the
file or the lint config changes.
Cancelled computations return `Err` and are **not** cached (via
`OnceCell::get_or_try_init`), so a cancelled lint/doc pass is recomputed
on the next request rather than caching a partial/empty result.
Legitimate empty results (parse errors, non-package-export files, etc.)
are cached.
### What is not cached
TypeScript and dependency diagnostics are intentionally left uncached,
since they depend on other files in the graph.
### Testing
Existing LSP lint/doc diagnostic tests pass (`lsp_lint*`,
`lsp_doc_lint_diagnostics`, `lsp_code_actions_ignore_lint`,
`lsp_diagnostics_*`).
Closes #13578
Closes denoland/divybot#469
Co-authored-by: divybot <divybot@users.noreply.github.com>
Co-authored-by: Divy Srivastava <me@littledivy.com>