Originally written by @HNR.

have int16 as a valid type.

to gameid, change fundamental sql request.

Add a new integration test testing the construction of the dataset from
the altorg and nle_data directories. These altorg games have been checked to see if they are
consecutive episodes, and the match up is correct.

In this mock directory we include extra mock ttyrecs that should not be
included in the dataset. Readmore about the mock dataset in altorg/about.txt

Also: change dataset to default shuffle games.

Remove all mention of torch, and instead just ship numpy arrays.

documentation of populate db.

We want to blacklist files that we have had trouble rendering in the
past. In experience this is only a handful of ttyrecs in millions.

NLE does not use the standard ttyrec format, used by alt.org and others.
It stores actions, as well as states, albeit in a different channel.
This commit introduces a version to the ttyrec format - original
ttyrecs are still 'ttyrec.bz2', whereas henceforth NLE generates
'ttyrec2.bz2'.

This will be more useful down the line, as readers can easily detect
which way to read ttyrecs.

This allows the dataset to know which version of ttyrec is being read,
which will allow the downstream Dataset object to chose the right way to read
the arrays in future.

This ttyrec version is always 1 for 'altorg' or old style ttyrecs, and
currently 2 for any nle generated ttyrecs.

Now the converter objects are aware of which version of ttyrec they are
decoding, have been passed it from the the dataset.

We will now start storing the in-game score in ttyrecs, to provide
greater information for offline learning methods. This is done by
utilising a different 'channel' for score. As it stands the channels
are:
 - 0 - terminal information
 - 1 - action input
 - 2 - in-game score (based on `botl_score()` via `blstats[NLE_BL_SCORE]`)

 Note the in-game score is recorded once, just before `nle_step` returns, at
 the same time that `obs->done` is set. This is done to avoid having to store
 the in-game score everytime the game flushes to screen, and to keep the final
 ttyrec size low.

Add support for v3 ttyrecs in the converter, and feed through to the
dataset. Also add support to the read_tty script.  Note the converter
now writes a score when the frame comes up, but only "flushes" the main
frame when the action comes in.

We test with ttyrec with an example of gaining core from picking up gold and
killing monsters. We expect the action at index t to correspond to the
state and reward just passed.