Add linear from unifold for a clean-path residual transformer

[hypnopump]

Transformer layers were not following clean-path residual principles (model starts at identity). This PR borrows the Linear module from UniFold and ensures all transformer residual layers start at 0 in training.

Bonus: modifies the random seed to affect all GPUs at the same time, preventing lack of reproducibility in multi-gpu setup

New models train faster now (blue is new, red is old):

[hypnopump]

Previous param update_freq was doing the job of gradient accumulation, but not dividing by the number of accumulated steps. Changed it so it behaves like gradient_accumulation and the effective_batch_size can be used with the same learning rate depending on the number of gpus and memory consumption.

To be fair, this could be a separate PR. Could implement it like this if it's wanted @guolinke

See expected behaviour from huggingface's accelerate:

Toghether with the multi-gpu seed, merging this PR will improve robustness of training params and the ability to dynamically adjust the effective batch size (num_gpus * batch_size * accumulation_steps) given compute availability