fix spurious metrics test failures

[davepacheco]

Fixes #2994. See #2994 (comment) for details.

I'm going to see if the instances test failure looks similar.

[smklein]

As mentioned in chats: This will work for "metrics on a single disk", but might not work as well for "system-wide resources", which are made during the populate step.

Should be fine for this test, but caused me issues with the instance metrics test, where the sample "cpu_count = 0" was taken before the test started.

[smklein]

Thank you for fixing this!

[davepacheco]

I confirmed that the test_instance_metrics test failure seen in CI looks the same:
#2994 (comment)

I updated the PR as discussed in chat so that we grab a timestamp as the very first thing when setting up a ControlPlaneTestContext. That's what we use now as the start time of our window when querying for metrics. The end time is Utc::now(), which as discussed seems like it should be fine now that we're waiting for oximeter collections to finish.

I was never able to reproduce the test_instance_metrics test failure locally so it'll be hard to be positive this fixed the problem, but it clearly fixes a problem and I will run a bunch of iterations and make sure I don't continue to see any of these failures.

[davepacheco]

This ran for a few hours without issue: while RUST_BACKTRACE=1 cargo test -p omicron-nexus -- integration_tests::instances integration_tests::disks; do :; done so I think this is at least an improvement.