packedsets: simplify implementation using Table

[alaviss]

Summary

Instead of using a custom ordered table implementation, convert packedsets to take advantage of Table and built-in set[T]. This cut down on the amount of code in the stdlib without compromising performance.

Details

This came out of the prior work on Roaring Bitmap, where I needed a few sample implementations to measure performance with. I found the original PackedSet resembles a simple ordered table with clustered lower bits, and experimented with building a simplified version based on existing stdlib components.

In this new structure, keys are grouped by their top N - 8 bits and indexed via a Table, with the bottom 8 bits packed into a set[uint8]. Due to the change in internal structure, items() no longer yield keys in insert order.

Performance metrics for nim c --hints:off --warnings:off -fc compiler/nim.nim (on the same compiler checkout):

Implementation	Compiler runtime	Ratio	Improvement
Original	10.469s ± 0.075s	1.0 ± 0.007	0 ± 0.007
Simplified	10.114s ± 0.107s	0.97 ± 0.011	0.03 ± 0.011

Implementation	Compiler peak memory	Ratio	Improvement
Original	1.137GiB	1.0	0
Simplified	1.123GiB	0.988	0.012

For the compiler, the new implementation performance is very similar to the original.

---

Notes for Reviewers

• While tests passes, the original suite isn't very comprehensive. Other compiler failures showed that the tests are not good enough. Draft until tests and microbenchs are written.
• I have not yet measured the performance of individual set operations in detail. The compiler mostly uses incl and contains while not employing many mathematical operations.