Fix rust_benchmark rule

[djmarcin]

I was just looking into fixing #231, but it doesn't seem like the current setup is well suited to criterion benchmarks (which I think has become the preferred method?). Specifically, when delegating to _rust_test_common, the criterion code doesn't seem to run at all because criterion wants to disable the test harness in favor of its own main.

Since this rule doesn't currently work, it's unlikely anyone depends on it. I propose we repurpose rust_benchmark to simply warn when the compilation mode != "opt" and delegate to a rust_binary with --bench which is more in line with how criterion expects to be invoked.

INFO: Build completed successfully, 60 total actions
Gnuplot not found, using plotters backend
fib 20                  time:   [36.696 us 37.153 us 37.602 us]
Found 2 outliers among 100 measurements (2.00%)
  1 (1.00%) high mild
  1 (1.00%) high severe

This strategy can optionally be made to work for the nightly #[bench] attribute by passing the --test flag to rustc to generate the test harness. This is enabled with the test_harness = True option.

A test was added using criterion to prevent future bit-rot.

[UebelAndre]

Hey, sorry about the long delay here. I'd be happy to take a look if you had the time to rebase 😄

[UebelAndre]

The changes here were merged in #923, thanks again!