Overhaul how the rustc benchmark works.

[nnethercote]

Currently the rustc benchmark is handled like a normal benchmark in
some ways, and handled very differently in others. This results in
various non-obvious special cases.

This commit improves things.

• Adds a --bench-rustc option to bench_local and bench_next, which
  makes the rustc benchmark opt-in.
• Omits the rustc benchmark from --include/--exclude handling.
• Calls it the "special rustc benchmark", in contrast to the "normal
  benchmarks".
• Handles it directly in bench(), avoiding a bunch of awkward
  special-casing code.
• Adds an eprintln! to make it more obvious when it starts running.
• Cleans up some unnecessary string conversions in bench().

[adamgemmell]

This seems like a good time to add a note in the README about how the artifact ID will be used as the rust-lang/rust ref to benchmark when using bench_local --bench-rustc. I don't believe it's mentioned anywhere in the documentation otherwise and it's quite important to know

[Mark-Simulacrum]

This looks largely good to me -- the note @adamgemmell suggests seems like a good idea.

This likely should also update https://github.com/rust-lang/rustc-perf/blob/master/collector/collect.sh#L29 to make sure we continue collecting rustc on the perf.rlo collector.

[nnethercote]

This likely should also update https://github.com/rust-lang/rustc-perf/blob/master/collector/collect.sh#L29 to make sure we continue collecting rustc on the perf.rlo collector.

It does: https://github.com/rust-lang/rustc-perf/pull/1140/files#diff-dcc00c21dcd21558f9180d8045820f2f8e5522e76a8880db41b6a89a31fa991a

[nnethercote]

I have fixed the formatting and added the note about the ID in the second commit.