automate or expose recovery of nonfunctional sleds

With two switches and multiple instances of nexus, it should be possible to distinguish a single (or small number of) nonfunctioning sled -- no connectivity with sled-agent -- from a network partition. While no partition detection algorithm is ever foolproof, we should expose to operators at least the option to automatically or manually recover a sled that is clearly not working. The most reliable means of automatic recovery with capture of debugging information, short of on-site manual intervention, is to induce an NMI via the SP, presumably by communication via MGS. While policy is a bit fraught here -- this is disastrous if fully automated and we incorrectly diagnose a network partition as multiple faulty sleds -- mechanism is not. There is already a mechanism for NMI generation from pilot to faux-mgs to hubris. A reasonable starting place would simply be an API call that is plumbed through, without ever being executed automatically.

Trace the existing NMI-generation path from pilot through faux-mgs to hubris, including how communication via MGS is represented. Define the proposed API boundary and verify that it can be invoked manually without automatic execution; done means the call is plumbed through and covered by the relevant existing tests.