precompute audio and text layers to be used for hierarchical contrast…

…ive loss

add the hierarchical contrastive loss from the ViCHA strategy in http…

…s:https://arxiv.org/abs/2208.13628

allow for sampling multiple pieces of music and selecting top match w…

…ith audio clip mulan

map location to cpu when loading state dict

fix save command not using the one on the trainer class, for saving o…

…ptimizer state

0.0.24

add qk rmsnorm

avoid a division

repeat batch for relative positional bias only if doing patch dropout

add best relative positional encoding for AST