[ty] Collect completions into a max-heap
The previous commits made this change deliciously simple. The key point
is that we don't collect a completion if it has no way of being ranked
into the top 1000 that we return.
Before:
```
$ ./target/profiling/ty_completion_bench ~/astral/relatedclones/scratch-home-assistant/homeassistant/scratch.py 1 -q --iters 30
total elapsed for initial completions request: 554.746713ms
total elapsed: 5.435318983s, time per completion request: 181.177299ms
After:
```
$ ./target/profiling/ty_completion_bench ~/astral/relatedclones/scratch-home-assistant/homeassistant/scratch.py 1 -q --iters 30
total elapsed for initial completions request: 526.743638ms
total elapsed: 4.268009725s, time per completion request: 142.26699ms
```
This is an especially nice speed-up for the cached case, where lots of
time (proportionally) was being spent sorting potentially a very large
list of completions.
There is potential future wins here. For example, every auto-import
completion has an `import` statement generated for it. We could instead
defer to that work until the very end when we convert the max-heap into
a `Vec`. Since before then, the completion might get dropped and there's
no point in doing extra work. But that will require more refactoring
(and possibly another `CompletionIntermediate` type? blech).
