VQVAETransformerInferer not compatible with a maximum sequence length

[marksgraham]

If we set max_seq_length to a value smaller than the full sequence length of the encoded image, this error is thrown upon training:

AssertionError: you are passing in a sequence length of 12585 but your absolute positional embedding has a max sequence length of 1000

Fixing this during training should be straightforward, if max_seq_length is smaller than the total sequene length we should train on a random section of the sequence of length max_seq_length.

I guess that we need to make get_likelihood more like the sampling function; so it loops and predicts the logits for the next token using a previous context of length max_seq_length-1.