Don't block the UI for prompt completions

[rhogenson]

The prompt handler for : commands currently blocks the UI while generating completions, which can make the editor feel laggy when using :o or :cd on networked filesystems. I was reproducing this issue using github.com/nirs/slowfs to inject a 1 second delay to readdir() calls (although in reality I experienced this issue while working in google's monorepo).

To see the difference, I recorded before and after videos of navigating a slowfs directory containing some nested directories.

Before the change: https://asciinema.org/a/yzdSZUjjnrBPAU1VcFcVAkLyu

After the change: https://asciinema.org/a/fgayEu1ZczVwqnMrAsmsvuPrJ

Fixes #11604

[the-mikedavis]

I have looked into this refactor in the past and couldn't find a way I really liked to solve it. We definitely want to move the I/O done for path completion off the main thread but the refactor to do so is not straightforward.

Refactoring Completer to drop the Editor arg and use jobs/callbacks has drawbacks. It's a very large change that affects nearly any place that we use completers (this part we may not be able to avoid). It's also pure overhead for all completers except path completion - fuzzy matching is so fast that doing it in the render loop is not worrisome. Plus it makes completers that need to interact with the Editor complicated/unnatural (for example buffer and language completers using a channel).

I chatted with @pascalkuthe about this in the past and he thought it might be a good idea to special-case the path completions or at least have a special case for any "deferred" completion values. As a straightforward way to do this we could have the Completer type return an enum that either is the full completion Vec<Completion> or a deferred value that would be debounced and calculated off the main thread in a hook.

[rhogenson]

Thank you! I was having trouble finding a good way to handle both synchronous and asynchronous completions, but that works really well! I think this change turned out a lot nicer, please take a look.