bazel: investigate performance impact of nogo

[rickystewart]

Compare cold and incremental build times between the following configurations:

1. Standard cockroach-short build
2. cockroach-short build with --config nonogo
3. Control: delete the nogo = ... argument from the go_register_toolchains() call in WORKSPACE.

For cold build times, we expect nogo will add a substantial initial startup cost before builds begin (this is case (1) above). This is because it needs to compile all the Analysis passes before building the actual Go binary, which does take a long time. (Unfortunately, because one of our passes depends on cockroachdb/errors which has proto dependencies, we need to compile all the C++ code as part of nogo.) However, you shouldn't have to repay that cost until the nogo config is updated. (2) and (3) should have similar overall build times for cold builds.

For incremental builds, (1), (2), and (3) should all be roughly similar. If (1) takes much longer than (2) or (3), or (2) takes much longer than (3), that is a problem.

Epic CRDB-8036

Jira issue: CRDB-11822

[irfansharif]

Another thing worth measuring: performance impact of actually running linters as part of dev build {short}, even if built with --config nogo.

[irfansharif]

Haven't investigated this fully/scientifically, but nogo has non-neglible overhead. I tried the following experiment on my GCE worker:

• dev build short (prime cache with nogo artifacts)
• dev build short -- --config=nonogo (prime cache with nonogo artifacts)
• Edit a random pkg/roachpb file to invalidate base package dependency
• Time dev build short vs. dev build short --config=nonogo for the incremental build

2:31:05s (nogo) vs. 1:51:02s (nonogo)

Tried this a few times when comparing make incremental builds vs. bazel when editing pkg/roachpb, was able to reliably see that nogo has overhead.

1:24.47s (make) vs. 2:31.40s (bazel,nogo)
1:24.95s (make) vs. 2:02.96s (bazel,nogo)

1:25:36s (make) vs. 1:42:62s (bazel, nonogo)
1:28:21s (make) vs. 1:45:65s (bazel, nonogo)
1:24:51s (make) vs. 1:46:09s (bazel, nonogo)

[irfansharif]

Earlier iteration of "should we default to building with nogo": #74540.