cleanup: remove dead MAGNeT/MBD loader code, drop redundant dbn=False, remove loopgen-repo reference copy

shorten comment

fix mask padding trim missing from non-memory-efficient attention path

route magnet through custom attention path instead of nn.MultiheadAttention

revert dead-code edit in loaders.py, real fix is in loopgen.py

disable memory-efficient attention in loopgen's own magnet loader

disable memory-efficient attention for magnet

try torch 2.10.0 to match xformers build, shrink gpu duration for testing

disable memory-efficient attention for musicgen too

disable memory-efficient attention, xformers has no blackwell kernel

initial deploy