Fix propagation of use_flash for offloaded inference

[epenning]

This change fixes an issue where _forward_offload() was being passed the parameter use_flash but didn't accept the parameter use_flash. This also passes use_flash along to _prep_blocks() since it's a required parameter.

Context

When running run_pretrained_openfold.py with the config for offload_inference enabled, this error occurred:

Traceback (most recent call last):
  File "/.../shared/openfold/run_pretrained_openfold.py", line 591, in <module>
    main(args)
  File "/.../shared/openfold/run_pretrained_openfold.py", line 439, in main
    out = run_model(model, processed_feature_dict, tag, args)
  File "/.../shared/openfold/run_pretrained_openfold.py", line 119, in run_model
    out = model(batch)
  File "/.../shared/openfold/lib/conda/envs/openfold_venv/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1130, in _call_impl
    return forward_call(*input, **kwargs)
  File "/.../shared/openfold/openfold/model/model.py", line 510, in forward
    _recycle=(num_iters > 1)
  File "/.../shared/openfold/openfold/model/model.py", line 384, in iteration
    _mask_trans=self.config._mask_trans,
TypeError: _forward_offload() got an unexpected keyword argument 'use_flash'

[gahdritz]

Thanks for the PR! This was actually half-deliberate; FlashAttention doesn't work that well for the long sequence sizes for which the offloading mode is useful. I propose completely removing the use_flash parameter being passed to _forward_offload in model.py and manually passing use_flash=False to _prep_blocks in the same function instead. I'll add a new constraint in the config preventing users from trying to combine the two. I changed my mind. I'll just add a warning to the config and let users do as they please.