rand_distr: Port benchmarks to Criterion

[vks]

• The benchmarks are now living in their own crate. Therefore, this does
  not add any dev-dependencies to rand_distr.
• Instead of bytes per seconds, we now measure cycles per byte.

Refs #1039.

[dhardy]

Nice!

Maybe benches should be a top-level crate instead of under rand_distr?

[dhardy]

This is how you're supposed to use it? $ cargo run --release

$ cargo +nightly bench
<snip>
exp/exp                 time:   [14380.2573 cycles 14551.8306 cycles 14750.6125 cycles]
                        thrpt:  [1.8438 cpb 1.8190 cpb 1.7975 cpb]
Found 1 outliers among 100 measurements (1.00%)
  1 (1.00%) high mild
exp/exp1_specialized    time:   [13641.4057 cycles 13718.3708 cycles 13800.0997 cycles]          
                        thrpt:  [1.7250 cpb 1.7148 cpb 1.7052 cpb]
Found 11 outliers among 100 measurements (11.00%)
  4 (4.00%) high mild
  7 (7.00%) high severe
...

[newpavlov]

Note that criterion is a quite heavy dependency, so I also think it's worth to keep these benchmarks in a separate crate. Otherwise, CI times may rise significantly.

[vks]

@dhardy Yes, this could be a top-level crate, but I wanted to get feedback before converting the other benchmarks. (There are top-level rand benchmarks already occupying the benches folder.)

As you already figured out, you can run them with cargo bench. Nightly is only required because of #![rustfmt::skip].

@newpavlov They are already in a separate crate, and not checked by CI. In fact, I implemented the benchmarks based on the way it was done in the RustCrypto crates. The next version of Criterion will make the HTML reports optional, which hopefully turns it into a lightweight dependency. If we are still worried about CI runtime, I think caching can solve this issue.

[dhardy]

Sorry, I guess I should approve this.

[vks]

Thanks!