Add `Dataset` and `Writer` modules for the `lmdb` format

[RasmusOrsoe]

Is your feature request related to a problem? Please describe.
lmdb is a memory-mapped database format that provides random access to individual training events faster than SQLite.

The main benefits of lmdb: Faster query speeds, lower storage footprint, serialized data allows for an elegant implementation of storing arbitrary DataRepresentations (see #781).

For larger datasets, and particularly large events, SQLite can become prohibitive as a dataset format.

Describe the solution you'd like

• A LMDBWriter that outputs events in lmdb format. Should accept a serialization method (msgpack, for example). Should optionally accept a DataRepresentation - if given, representations are pre-computed and serialized using dill or similar pickle methods (see Graph construction before training #781). A field in the database should contain relevant information regarding the serializer, such that the file is a self-contained object that users can read without prior knowledge of the serializer used.
• A LMDBDataset that is compatible with the lmdb database format. Should automatically check for which serializer was used, so the user doesn't have to guess. Should be able to retrieve pre-computed data representations.

[astrojarred]

I have a working version of LMDBDataset. If it would be useful I can work on building up a writer as well!

[RasmusOrsoe]

That would be great! I'll assign you to the issue!