not all the transformer may need conditioning, perhaps only semantic …

…is conditioned, the rest just upresolutes

get a working example of text conditioned semantic transformer training

switch to modrelu, seems to be recommended by a lot of torch complex …

…packages

allow fine transformer to be trained on pre-encoded soundstream codeb…

…ook indices

set things up for eventual text conditioned audio synthesis

expose the ema decay for the residual vq in soundstream

some tweaks to soundstream, include commitment loss

validate transformers being passed into audiolm