G3NetworkSender background serialization

[cnweaver]

Refactor G3NetworkSender so that it can handle larger rates of large frames, for which frame serialization overhead can dominate over transmission time. To do this, the option is provided to run some number of background threads responsible for serializing frames. The default is to use zero serializer threads, and to continue to serialize on the main thread, as has previously been the case. Because the access to the serialized blobs inside G3Frame is not currently thread safe, frames are not emitted from G3NetworkSender until their serialization is completed. The mechanism to coordinate this also adds some small overhead to the case of serializing on the main thread, which appears to be unimportant, but which could be avoided if there is a need to do so.

[cnweaver]

There's a potential deadlock in this code, so it should not be merged just yet. I'm working on fixing it. (Roughly the same failure mode existed in the existing version, but was exceedingly difficult to trigger; unfortunately, by adding complexity I increased expose to it.)

[arahlin]

I think this is ready for review? Some minor nits below, otherwise looks good, assuming it works.

[cnweaver]

Yes, I think this is ready for full review; the gotchas that I found during testing or which nwhitehorn warned about should now be fixed. I've given this a fair amount of testing myself, particularly in the multiple serialization thread configuration, and also some more artificial testing with serialization on the main thread, and I have observed no further issues. However, if there are any particular tests which seem like they should be tried, let me know and I'll see what I can do.