Fix HashSet::union performance

[stepancheg]

Consider this example: small_set = 0..2, large_set = 0..1000.

To efficiently compute the union of these sets, we should

• take all elements of the larger set
• for each element of the smaller set check it is not in the larger set

This is exactly what this commit does.

This particular optimization was implemented a year ago, but the
author mistaken < and >.

[rust-highfive]

r? @alexcrichton

(rust_highfive has picked a reviewer for you, use r? to override)

[alexcrichton]

@bors: r+ rollup

[bors]

📌 Commit 04a237b has been approved by alexcrichton

[ssomers]

What happened is that I measured it on free standing code with free standing benchmarks, and the code must have gotten mangled while transplanting it into libstd, and I didn't even try to transplant benchmarks. In general, there are few benchmarks in libstd/collections/hash compared to liballoc/benches/btree. I could make some but I guess it's too late to include in this PR.

[ssomers]

cargo benchcmp of a few benchmarks on modestly sized sets (10 versus 100 elements) says about this change:

name                                                before ns/iter  after ns/iter  diff ns/iter   diff %  speedup
collections::hash::bench::set_difference            124             125                       1    0.81%   x 0.99
collections::hash::bench::set_intersection          122             126                       4    3.28%   x 0.97
collections::hash::bench::set_symmetric_difference  1,624           1,821                   197   12.13%   x 0.89
collections::hash::bench::set_union                 1,581           266                  -1,315  -83.18%   x 5.94

Quite a lot better than correcting the "still measurable" improvement I reported in #57043, whatever it was that I tested.