sled-agent could handle Unix signals better

[dancrossnyc]

Right now, they seem to be deliverable on any thread, which can interrupt all sorts of things, sometimes with deleterious effect (cf, stlouis#922).

The most likely source of signals is probably when it spawns a (Unix) processes that then exits and a SIGCHLD is delivered to sled-agent, which sled-agent does with some non-trivial frequency. As a daemon, sled-agent could probably handle this better; perhaps disabling signals on all but one (dedicated) thread that blocks in sigwait or something similar.

[davepacheco]

We discussed this at some length today and came up with this: oxidecomputer/oxide-tokio-rt#3

[hawkw]

oxidecomputer/oxide-tokio-rt#4 adds the requisite API to oxide-tokio-rt to allow spawning a dedicated signal-handling thread. I'd love reviews on that. Once it's merged, I'll release a new oxide-tokio-rt and come back to make sled-agent use the new API.