fix: support opengptx data encoding

[lhahn-iis]

When playing around with the Dataset implementations of PackedMemMapDataset using OpenGPT-X data, decoding errors happened. This was also discovered in the past, that this data cannot get properly decoded using "utf8".
Although this very likely to be problem with the conversion of the respective data and the resulting encoding there, a potential solution is provided within this PR.

Note: This branch is based on #9 in order to use the respective test infrastructure and should get therefore merged afterwards, but not beforehand.

[lhahn-iis]

It is worth noting that this problem can also get solved by simplifying the general way of accessing files here. Right now we are basing the functionalities on np.memmap. We could however also use IOStreams with a certain offset and a specific read()-length. Some small attempts to play around with that indicate also much better performance (which makes it also interesting for #23. I decided however not to add it for now, since these changes might collide with the future use of more than text data.
If we want to add this anyhow, some more time needs to get invested to verify the stability of the code, since there seem to be still some problems with this implementation.

We can however also leave this to #23 and keep the changes here as a quick solution to support OpenGPT-X Data.

[lhahn-iis]

Closing this, see #40 for more details