Use binary search in ArrayRangeSet

[fabalchemy]

This PR modifies the ArrayRangeSet implementation to use binary search instead of linear scan for insertions and removals, as suggested on Discord.

I also wrote two basic criterion benchmarks:

• completely random insertions and removals between 0 and size
• almost-linear insertions between 0 and size with occasional holes

But did not include them in this PR since I'm not sure they're super meaningful and require a bunch of noisy diffs in quinn-proto to expose the types for the benchmarks.

They show limited gains (but no regression) for smaller sets, but significant improvements for larger sets, especially for the linear scenario which I believe is closer to the expected usage in quinn.

Size	Mode	BTree	Array (linear scan)	Array (binary search)
100	Random	14.431 µs	4.7897 µs	4.5871 µs
100	Linear	13.288 µs	4.9711 µs	4.2885 µs
500	Random	95.934 µs	46.976 µs	25.327 µs
500	Linear	84.387 µs	56.173 µs	22.717 µs
5000	Random	1.2510 ms	3.0309 ms	378.20 µs
5000	Linear	1.0457 ms	3.8295 ms	256.83 µs
10000	Random	2.6267 ms	11.886 ms	1.0308 ms
10000	Linear	2.1992 ms	13.944 ms	533.71 µs

[Ralith]

Nice use of partition_point!

Does this have a statistically significant impact on e2e benchmarks? The microbenchmark results look great, but I want to be sure we're not pessimizing real-world workloads, which may not be fairly represented by bulk inserts.

[fabalchemy]

I ran the quinn benchmarks on my machine and it seems this change has no impact.

I'm a bit surprised by the noise, but this is expected judging by other PRs.

Linear scan:

test large_data_10_streams  ... bench:  20,074,834 ns/iter (+/- 7,122,084) = 522 MB/s
test large_data_1_stream    ... bench:   1,995,283 ns/iter (+/- 3,274,522) = 525 MB/s
test small_data_100_streams ... bench:  13,939,785 ns/iter (+/- 30,021,101)
test small_data_1_stream    ... bench:  28,819,057 ns/iter (+/- 11,289,925)

Binary search:

test large_data_10_streams  ... bench:  19,686,293 ns/iter (+/- 5,402,187) = 532 MB/s
test large_data_1_stream    ... bench:   1,948,904 ns/iter (+/- 3,207,186) = 538 MB/s
test small_data_100_streams ... bench:  14,169,080 ns/iter (+/- 37,016,181)
test small_data_1_stream    ... bench:  29,511,801 ns/iter (+/- 10,850,156)