Add ability to deep-mind log to stderr

[heifner]

The theory is that #724 is caused by too much buffered output to the pipe. Add ability to write to stderr which is unbuffered.

Add an example logging.json that can be used to configure deep-mind to log to stderr.

Seems to resolve #724

[spoonincode]

Adding the ability for deepmind to go to stderr seems useful but also seems like a new feature (i.e. not release branch material).

So far I've not seen any evidence that a backed up FILE buffer can cause this failure. From the testing I've done, when the buffer is full nodeos blocks as expected. Is there a way to reproduce this error just from buffering? The EBADF errno from the original report also would not immediately suggest a user-space buffering problem to me.

Since deepmind output is, I take it, always expected to be piped to some other process, maybe dmlog_appender should fflush after each output (or maybe more optimally at some other delineation point). That does seems like a reasonable thing we could add. But lack of that only means that worst case the consumer is a little "behind" the most recent deepmind output.

imo we need more understanding of the issue -- we can't just say deepmind only works properly on stderr instead of stdout.

[heifner]

@spoonincode re-targeted to main.

[ericpassmore]

Note:start
category: Logging
component: DeepMind
summary: Add the ability for deepmind to go to stderr, which is unbuffered.
Note:end