Some tor-proto tests seem unreliable now

This test in tor-proto seem to be unreliable since we moved from tokio::test to test_with_all_runtimes in #222 (closed), especially when building with --all-features. Here is a sample:

---- channel::reactor::test::new_circ_create_failure stdout ----
thread 'channel::reactor::test::new_circ_create_failure' panicked at 'assertion failed: matches!(cell_sent . msg(), msg :: ChanMsg :: CreateFast(_))', crates/tor-proto/src/channel/reactor.rs:535:13
note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace
failures:
    channel::reactor::test::new_circ_create_failure

If I add a dbg! to the test so I can see the value of cell_sent.msg(), I get:

[crates/tor-proto/src/channel/reactor.rs:535] &cell_sent.msg() = Destroy(
    Destroy {
        reason: DestroyReason(NONE),
    },
)

I'm guessing that maybe we're dropping a reactor before we should? That might cause a Destroy cell to get sent when we're expecting a CreateFast.