Migration ought to consider block backends more generally

this is an "I read the code and see no way it works"-tier issue. reproduction is probably(?) straightforward but I tripped over https://github.com/oxidecomputer/propolis/issues/981 along the way.

in propolis-server we add a devices entry for Blob backends. the comment there is sound: we should migrate writes to the blob otherwise they'll get lost in migration. but if there is block-local state for other backends (say, a collection of coalesced-but-not-yet-delivered I/Os) we need to remember to plumb up that block backend too or lose the I/Os in migration.

propolis-standalone has a worse time of it: we look at devs, but not block, so I'm pretty sure we'll lose mem-async backend contents across migration (this is what I wanted to prove to myself before tripping over #981)

I've been kind of sort of working on an I/O coalescing patch that would bring these problems and more to file backends, and I would absolutely have forgotten to plumb up the block-side state migration at first :(

Start by reading the Blob device setup in bin/propolis-server/src/lib/initializer.rs and the device handling in bin/propolis-standalone/src/snapshot.rs. Reproduce migration with a block backend such as mem-async, then trace how block-local state is selected for migration. Done means block backends are handled generally so their pending contents or I/O state is not lost.