Limit memory usage for captured logs or stdout

I've recently spent some time debugging a test suite that started OOM'ing. The culprit was a runaway greenlet which logged as fast as it could, but due to caplog's disposition to store logs until test's end, the memory leak was well obscured from me. It was even stranger since due to its nature (a runaway greenlet), I couldn't attribute it to any particular greenlet, yet the longer the test suite ran, the more memory it leaked.

Maybe we could implement practical per-test and per-suite log capturing limits, which could be bumped through configuration, but otherwise would be helpful for cases such as these?

Start by tracing pytest's caplog and stdout capture entry points to understand where output is retained for a test and for the suite. Define practical configurable limits and determine what behavior should occur when they are reached; done means runaway captured output cannot cause unbounded memory growth while normal test capture remains useful.