Fix pipeline's context overflow

[maxreciprocate]

This PR

• truncates input into pipeline to bert's context length (512)
• puts pipeline on processes' local gpus for 2x speedup in training on single gpu (1:02 v 2:19 for 128 gradient steps), a bit more on multigpu (1:07 v 9:26 on 8 gpus) at cost of 0.5G extra VRAM
• removes depreciation warning of return_all_scores usage in pipeline, on transformers version 4.24.0