Cache step values, saner logging levels

Breaking changes

There are no breaking changes.

New features

We now support a form of caching events traced at every step through the rlog.ValueMetric.
You can now configure it like this:

    ...
    rlog.ValueMetric("mse_err", metargs=["err"]),
    ...

and in your hot loop do:

    train_log.put(err=err)
    if logging_time:
        rlog.trace(step=policy.steps, **rlog.summary())

Improvements

The logging levels are now:

Level	Numeric value
CRITICAL	50
ERROR	40
WARNING	30
INFO	20
TRACE	15
DEBUG	10
NOTSET	0

• RLog is now configured with TRACE as before but
• the StreamHandler writing to the console and FileHandler writing to the the text file are configured as INFO in order to avoid dumping the data buffered by rlog.ValueMetric
• the structured handlers such as PickleHandler and TensorboardHandler are configured to TRACE levels and are logging this data to disk.