`tach server` runs `check` too often

[emdoyle]

Right now, the LSP server built into Tach is very naive. When a Python document is opened or saved, it will run a full tach check and then filter the errors to the ones relevant to the file.

A relatively simple improvement would be to create an API for checking a single file at a time, since this should be fast enough (even for large files) to keep the current eager behavior.

Another avenue would be to make the server only re-check the project when Python files or the tach.toml are saved. When a file is opened, it could read the check results from a cached copy in memory, and on change it could either do nothing, or publish empty diagnostics. This seems more complex though, and probably doesn't perform noticeably better in most cases.