updt setup.py; fix tokens_per_expert casting

[vchiley]

PR does a few little things:

• if setup.py imports torch therefore torch should already be installed; it shouldn't be part of the script install_requires.
  • torch similarly removed from stanford-stk @ git+https://github.com/vchiley/stk.git@setup_deps
• use torch.cuda.get_device_capability() to set --generate-code compile flag for nvcc
• autocast before all2all if autocast enabled
• fix casting issue in def load_balancing_loss(*args)
• create megablocks/backend/__init__.py