Add a `paddle usage` command to support collecting demo usage.

Fix #1008

* Make this demo support GPU

Once encoutering a single sample whose size is larger than batch size,
the provider will yield empty batch and terminate the current pass unexpectedly
if can_over_batch_size=False.

* Partially fix #622

In order to keep consistent with other evaluators, an extra parameter
label is add to chunk_evaluator. Document and demos are modified
accordingly.