Bound the length of nls completions

[jneem]

When there's a big contract, the completions can be very large. This can slow down nls substantially, with most of the time being spent in pretty-printing. The current PrettyPrintCap mechanism isn't useful here, because it does all the slow pretty-printing first and then truncates the output.

This PR adds a crude mechanism to bound the size of pretty prints. With this change, completion in the tf-ncl example is instant. There are probably other places that truncated pretty-printing could be used in nls.

I'm not completely happy with how this implementation threads the bounds through, but I couldn't figure out how else to do it:

• I could use a wrapper type like BoundedTerm<T: Pretty>(T, SizeBounds) to wrap the things that are actually getting pretty-printed, but it requires lots of changes to the pretty-printer to explicitly pass that state around everywhere
• I tried keeping the state in the allocator and recursing with smaller "child" allocators, but the lifetimes in DocAllocator don't allow this: since the lifetime of the allocator leaks into the returned DocBuilder, you can't format children using a local allocator.

The current approach with interior mutability was from me trying to work around the second issue.

I also wonder if NickelAllocatorExt needs to be a trait. Maybe we can just use BoundedAllocator (possibly renamed) concretely everywhere?

[github-actions]

Bencher Report

Branch	2073/merge
Testbed	ubuntu-latest

⚠️ WARNING: The following Measure does not have a Threshold. Without a Threshold, no Alerts will ever be generated!

• Latency

Click here to create a new Threshold
For more information, see the Threshold documentation.
To only post results if a Threshold exists, set the --ci-only-thresholds CLI flag.

Click to view all benchmark results

Benchmark	Latency	nanoseconds (ns)
fibonacci 10	📈 view plot ⚠️ NO THRESHOLD	487,290.00
foldl arrays 50	📈 view plot ⚠️ NO THRESHOLD	1,768,000.00
foldl arrays 500	📈 view plot ⚠️ NO THRESHOLD	6,859,700.00
foldr strings 50	📈 view plot ⚠️ NO THRESHOLD	7,212,800.00
foldr strings 500	📈 view plot ⚠️ NO THRESHOLD	62,931,000.00
generate normal 250	📈 view plot ⚠️ NO THRESHOLD	48,366,000.00
generate normal 50	📈 view plot ⚠️ NO THRESHOLD	2,135,300.00
generate normal unchecked 1000	📈 view plot ⚠️ NO THRESHOLD	3,373,100.00
generate normal unchecked 200	📈 view plot ⚠️ NO THRESHOLD	770,260.00
pidigits 100	📈 view plot ⚠️ NO THRESHOLD	3,200,300.00
pipe normal 20	📈 view plot ⚠️ NO THRESHOLD	1,512,700.00
pipe normal 200	📈 view plot ⚠️ NO THRESHOLD	10,392,000.00
product 30	📈 view plot ⚠️ NO THRESHOLD	838,970.00
scalar 10	📈 view plot ⚠️ NO THRESHOLD	1,542,500.00
sum 30	📈 view plot ⚠️ NO THRESHOLD	842,330.00

🐰 View full continuous benchmarking report in Bencher

[yannham]

I agree that I don't see a much better way to do it, without butchering the actual pretty-printing code. I also think that the NickelAllocatorExt becomes useless - it's was done only because we wanted to monkey-patch the external type BoxAllocator, but now we can just implement the methods directly on our own type.

[jneem]

Ok, the new version gets rid of the NickelAllocatorExt trait and a bunch of generics. It also uses a closure for scoping and avoids underflows.

I couldn't figure out the purpose of the A parameter. Even within pretty, it just seems to get passed around without doing anything. So I got rid of it also.