[WIP] FedGlobalContrast and FedSimCLR baseline

[xkxxfyf]

current problem: global calculate contrast loss is slow because of epoch data size, but it will take more communication cost with batch size calculate global contrast loss

[joneswong]

provide a unit test (see test folder) for basic fed contrastive learning and fine-tuning.

[xkxxfyf]

the revise of keeping the same Non-IID distribution of train and test has been verificated by fedsimCLR pre-train weight and local linear-prob evaluation. The client avg acc is 0.723 for val and 0.735 for test. It close to the global acc of 0.715 for val and 0.693 for test

[rayrayraykk]

Please add unit-test ASAP, thanks.

[joneswong]

we have suggested to add UT, but there is still nothing

[rayrayraykk]

1. It is ok to alternatively make local and global updates. However, we have suggested to enable a combination of these losses in one gradient descent step, but it seems that this implementation still fails to do that. 2. There are still many issues remaining. 3. This pr would not pass the check of a linter, imo, right? @rayrayraykk .

The quality of this pr is exceptionally low. Please complete it ASAP.

Yes, the unit-test provided still does not work.