Logs can be dropped when threaded

Due to the quick fix for #128 (closed), we can lose individual logs in threaded situations.

I propose either:

1. figuring out some easy cross-platform locking to build a queue in C;
2. having a ring buffer of log messages.

In my experience playing with this, I never saw more than 20 messages queued up. Making a 32kB ring buffer would solve that (albeit smell bad).