* NCCL based 1-bit Implementation + Refactor to add communication backends (#593)

* add nccl 1-bit optim.

* temporary commit to save stuff.

* Use dist collectives instead of mpi routines.

* remove old code for comm.

* Fix bugs. still does not work.

* modify to test the nccl side code path

* Initial gather impl. Works intra-node.

* Updates to comm. phase 2. nccl comm. passed the tests.

* refactor code to introduce nccl/mpi as backends for onebit adam.

* Refactor updates to test/engine.

* Fix compile/runtime errors.

* simplify support for nccl/mpi backends.

* Add missign file

* Add compression backend in constructor. Revert later.

* modify test with some perf counting.

* Implement a true non-blocking gather for nccl side.

* Revert "Add compression backend in constructor. Revert later."

This reverts commit df8c40d3.

* improve the 1-bit adam test.

* Refactor comm. and compression backend in 1-bit adam.

* Fix the test.

* Fix runtime errors and typos in nccl backend

* fix mpi backend. modify tests.

* modify nccl perf test.

* fix mpi side errors.

* Add an mpi perf test

* Sync DSE.

* Remove old collectives file.

* Undo a typo.

* Graceful failure for torch versions that don't support nccl pt2pt.

* Switch from deprecated allreduce interface.

* Make pipeline checkpoint files portable.

RTX-30 series are compute_86
```
python -c "import torch; print(torch.cuda.get_device_capability())"
```
This PR adds support for this compute capability.

Reference: https://developer.nvidia.com/cuda-gpusCo-authored-by: NJeff Rasley <jerasley@microsoft.com>

* 1) Register layout as buffer of module so that we can save/load checkpoint; 2) Add a broadcast of layout at the beginning to ensure different processes will have consistent layout during distributed training.

* Add docstring for max_seq_length argument in SparseSelfAttention
Co-authored-by: NZhun Liu <zhunliu@microsoft.com>
Co-authored-by: NJeff Rasley <jerasley@microsoft.com>

* [cifar tutorial] improve readability 

* tracking optimizer step in cpu-adam when loading checkpoint

* add warning/error message for updating optimizer step count

* resolve build issue

* supporting state update from the python side

* track step from python in all cases

* remove comma

* supporting different hidden dimensions

* add support for larger hidden dimensions (greater than 8K)

* remove empty line

* add loop unrolling factor for dropout kernels

* update different kernels based on the reviews
Co-authored-by: NJeff Rasley <jerasley@microsoft.com>

This PR:
* fixes a misspelled method name
* also `( () )` doesn't read too well, until one reads the code and understands that it's not a formatting bug. I proposed to simply say that it's a callable object.

Co-authored-by: NJeff Rasley <jerasley@microsoft.com>

Co-authored-by: NJeff Rasley <jerasley@microsoft.com>

In the absence of a model parallel group, model_parallel_allreduce should not do any reduction. This commit fixes the bug which was doing a model parallel allreduce across world group when model parallel group is None

* Use zero-tensors for missing gradients to avoid size mismatch

* Unit test for unbalanced gradients in ZeRO

* Formatting fixes

* Use AML method to set env vars instead of using mpi4py.
Co-authored-by: NJeff Rasley <jerasley@microsoft.com>

* Add guard to not using `torch.version.cuda` above no-CUDA environment.
* Fix several typos on setup.py.
Signed-off-by: NSeunghwan Hong <seunghwan@scatterlab.co.kr>
Co-authored-by: NJeff Rasley <jerasley@microsoft.com>

* zero-1 memory fix

* auto-tune max elems per comm to reduce padding/comm intervals

* clean-up and added previously missing reduction options

* fix testing backing to work with torch1.7