[ChunkTeacher] Fix infinite padding output

[emilydinan]

Patch description
Returning None here causes an issue where only padding examples are loaded onto the sample queue until we hit a crash. This was most obvious during dynamic batching because a self_observe invariant error was triggered.

Testing steps

 parlai tm -m test_agents/unigram --dict-tokenizer gpt2 -bs 2 -dynb full --truncate 128 -t integration_tests:chunky -mf /tmp/foo -dt train:stream