feat: working var3 training

fix: fix loss mask

fix: fix loss in var3 model

feat: add speech-flow based model

feat: variant 2 inference

wip: working on var2 notebook

wip: working on evaluation

fix: embed time as token, fix skip connections implementation

wip: working on second variant model

wip: working on e2 style

fix: fix attention input shapes

fix: fix shape

ref: remove seamless

ref: remove seamless

feat: working training