(I ran into this while shaving yaks as a part of #9832.)

There are three changes here:

af1090c changes Nexus::mg_clients() to look up MGD in DNS to find ports instead of assuming the hardcoded MGD_PORT. This allows mg_clients() to be useful in tests. (#[nexus_test] assigns random ports to services, but does put those ports in DNS.) This is kinda grody; we end up doing three DNS lookups (dendrite, MGS, and MGD) to build a HashMap<SwitchLocation, MgdClient>), but I didn't want to try to tackle that in this same PR. #8999 is related, and discussions some possible options that would let us squash this down.

ae54bbc adds a very simple test that hits the BFD status endpoint, expecting it to be empty. Without the fix from the next commit, this test fails with a 404 because #[nexus_test] only sets up one scrimlet (switch0), causing the BFD status endpoint to 404 because it can't find the other switch:

16:39:39.882Z INFO 913233fe-92a8-4635-9572-183f495429c4 (dropshot_external): request completed
    error_message_external = not found: switch with name "switch1"
    error_message_internal = not found: switch with name "switch1"
    latency_us = 385976
    local_addr = 127.0.0.1:38549
    method = GET
    remote_addr = 127.0.0.1:43174
    req_id = 262916c6-77d8-439f-8252-cede9b46f169
    response_code = 404
    uri = /v1/system/networking/bfd-status

bd79324 changes bfd_status() to skip missing switches instead of 404ing, so we still get status from the one switch. This allows the test in the previous commit to pass.