This patch effectively reverts cockroachdb#45983. The leaktest will now re-panic
when invoked on a panic unwind, instead of failing the test but
otherwise swallowing the panic and letting further tests run. Recovering
from panics is generally not a Go thing to do. The leaktest can only
recover from panics on the test's main goroutine (panics on other
goroutines, for example panics in a TestServer/TestCluster still crash
the binary), so the leaktest's behavior is inconsistent at best. The
panic-swallowing behavior is also surprising and magical - tests using
leaktest get it, tests not using leaktests don't get it, and figuring
out this difference is hard. Additionally, having the leaktest recover
panics is contageous - it puts pressure on other random code to be
mindful of its behavior under panic unwind. For example, the Stopper has
complex logic for recovering panics and doing best-effort cleanup,
recovering more panics from confused cleanup code, so that ultimately
future tests can run in a more or less sane environment.

I don't think the magic is worth it. Let's try again without it.

I believe that, at the time the leaktest magic was added, panics in go
tests were not resulting in a very clean experience in CI - it wasn't
trivial to see what test failed. I don't know if that's still the case.

Release note: None
Epic: None