Four sparsity layouts from DeepSpeed are now available for blocksparse attention on xFormer:
Fixed
BSLongformer
BigBird
Variable
sparsity_configs.py (https://fburl.com/code/s2n7x8gs) contains flexible objects with many parameters. The default parameters can be invoked through the quick_ functionality in attention_patterns.py (https://fburl.com/code/hya0t9e7) the produced layouts can be turned into a pattern with layout_to_pattern function (https://fburl.com/code/1qmsntyj)
Detailed notes for the task: https://docs.google.com/document/d/1cBlZeccvphI-d5avLkgKwZ4ScXQM1q6igqvlyFDXdDc/edit?usp=sharing

linting errors fixed for std-blocksparsity-attention-layouts

Added copyright
fixed block -> block_size
removed DeepSpeed TODO