Media support

[lgessler]

Users often derive their texts from an audio or video recording, and we want to be able to have these inside of each document. This requires that we (1) provide support for uploading and viewing media, and (2) extend the data model to support time-alignments.

The former should be straightforward enough, but the latter will take some thought. In ELAN, typically something like a sentence is transcribed first, and then tokens are created on a child tier. To have complete parity, this would seem to force us to have time alignments on both tokens and the "sentence" (cf. #41). The redundancy might be fine, but the issue is that it's unclear what the structure for the "utterance" ought to be. This issue and #41 should be solved in close consultation with each other.