As per https://gitlab.com/gitlab-org/gitlab/-/issues/321511 we've seen
one of these tests fail. It seems likely the problem is caused by the
use of `X.days.ago` throughout and then filtering with date ranges. The
likely explanation is that when the test is running just before midnight
then `2.days.ago` actually ends up being "3 days ago" from a date
perspective by the time the assertion is made about it.

Even though it's not easy to replicate this exact failure mode and this
flakiness has never happened for this spec before (to my knowledge) in 3
years it still is not a good way to write these kinds of tests as this
is theoretically a problem.

Instead of making all the times relative to "now" (which changes) I've
just made everything relative to some arbitrary fixed time. This works
fine, without any need for freezing time, as none of the methods being
tested actually refer to the current time.

The test is still quite difficult to read due to many different dates
and times and the fact that the examples are separated from the test
setup but I don't think this change makes it less clear.