[rllib] Document auto-concat in a3c

[ericl]

No description provided.

[AmplabJenkins]

Test FAILed.
Refer to this link for build results (access rights to CI server needed):
https://amplab.cs.berkeley.edu/jenkins//job/Ray-PRB/7086/
Test FAILed.

[AmplabJenkins]

Test FAILed.
Refer to this link for build results (access rights to CI server needed):
https://amplab.cs.berkeley.edu/jenkins//job/Ray-PRB/7094/
Test FAILed.

[ericl]

So based on tests, this actually degrades async mode more than expected, because what this auto-concat is doing is reducing the amount of off-policy data.

I'm pretty sure we can get sync sampling to work as fast as async with some batch size tuning, but for now, just document it so we don't need to worry about perf regressions.

[AmplabJenkins]

Test FAILed.
Refer to this link for build results (access rights to CI server needed):
https://amplab.cs.berkeley.edu/jenkins//job/Ray-PRB/7116/
Test FAILed.