Migrate hover, signature help, and completions to ArkFile

[lionel-]

Branched from #1263
Progress towards #1212

The handlers migrated use a slightly different approach because they use r_task() heavily and the Oak DB shouldn't be touched from an r_task() because of the risk of cancellation panics, which would be UB over the R stack. We could catch panics in r_task() and unwind for cancellation but I didn't want to think through the various implications.

• DocumentContext drops Document and stores the pieces it needs from the ArkDb and LSP state (tree, contents, line_index, encoding). This way the DB is not sent over an r_task().

• Add LSP <> TS position/range converters as methods on DocumentContext (parallel to the ones on ArkFile, also replacements for the Document methods).

• Migrate hover, signature help, and completion to the refactored DocumentContext.

[DavisVaughan]

Looks good, my main comments are about reducing duplication and noise

[DavisVaughan]

Should it be renamed to ArkFileContext at some point?

[DavisVaughan]

I would like to point out that we have

    // TODO(oak/completions): Clone so the closure captures by value. `r_task()`
    // sends the closure across threads, and `&WorldState` isn't `Send` because
    // `OakDatabase`'s salsa storage keeps thread-local query state.
    let state = state.clone();
    let completions = r_task(move || provide_completions(&context, &state))?;

So is something wrong here?

It's pretty odd to me that we:

• Are creating DocumentContext to make context here because of ArkDb issues
• Then we send over WorldState, which has those exact issues

Something feels off

[DavisVaughan]

Like, isn't this us using a salsa db inside an r_task? Which we definitely aren't supposed to do in case it panics?

[lionel-]

The completion handler runs on the main loop and thus can't be cancelled.

That said I agree this state of things is not great, but it's actually "fixed" in this PR: #1269

[DavisVaughan]

completion handler runs on the main loop and thus can't be cancelled

ah because it blocks the next request/notification handler from running at all, so we can't process a did_change() that might result in us being cancelled. gotcha.