[scale] Implement activation checkpointing for transformers.

[stephenroller]

Patch description
Leverage Fairscale to add in activation checkpointing. Activation checkpointing works by throwing away the activations during the forward pass, and recomputing them just-in-time during the backward pass, thereby trading away compute for memory savings.

It is particularly powerful when combined with FSDP. Utilized with --checkpoint-activations true

Testing steps
New CI test to just ensure that the path doesn't crash. Due to the simplicity of the call, I rely on fairscale's tests for correctness.

Manual testing: parlai train -m bart -t taskmaster2 -bs 4 -dynb full -tstep 100 --checkpoint-activations true. Took some very non-scientific measurements but were exactly as expected:

Checkpoint	Memory (nvidia-smi)	Updates/sec
No (baseline)	10823mb	1.01 u/s
Yes	7853mb	.77 u/s

[meganung]

lgtm! a clean wrapper add!