Add num_proc to map and filter calls

[dakinggg]

Adds the num_proc arg to the calls to huggingface dataset map/filter.

before: num-proc-test-baseline-1-cDb1Vk
after: num-proc-test-1-4nqb1g

dataset map goes from ~27 seconds to ~4 seconds for dolly_hhrlhf. Should be a bigger win for bigger datasets.

[dakinggg]

@alextrott16 I vaguely recall you had an issue with adding this at one point. Do you remember this? Seems to work fine for me now.

[irenedea]

LGTM! question and super nit

[alextrott16]

LGTM. Thanks for doing this.

@dakinggg I don't have a particularly clear memory of running into any issues with using num_proc. If you feel confident that our test coverage would catch any issues resulting from this leading to wonkiness, then I'm satisfied.