Close the performance gap for available strategy

Currently, we still face a performance gap for some of the existing strategies.
The expected performance can be found in the comments of the related `experiments` folder.

Any help in closing the gap is welcome. Just comment this issue and I will assign you to that strategy.

List of strategies to "fix":

- [ ] Elastic Weight Consolidation on Permuted MNIST
- [ ] Synaptic Intelligence on Permuted MNIST
- [x] iCarl on Split CIFAR-100
- [x] RWalk on Split MNIST
- [ ] GSS on Split MNIST
- [x] COPE on Split MNIST (most likely a bug on COPE)
- [x] LaMAML on Split Tiny-ImageNet