Eval harness fix

[sdtblck]

This fixes some problems relating to eval harness & model parallel models.
To summarize:

• logits weren't being gathered automatically, so with a model parallel size of 2, each machine only had vocab_size / 2 logits, which resulted in indexing errors.
• There was no straightforward way to toggle whether the logits were gathered or not, now there is (model._set_parallel_output(value))
• When zero optimization was being used in a config, it would break at inference time. Fixed this by overriding the zero config when setting up for inference (https://github.com/EleutherAI/gpt-neox/compare/eval_harness_fix?expand=1#diff-b6ee2c1db780b46787ab3a576020f9a12695f3f5cb38ba8badcd5501960e5d22R409)
• There was no way to specify which tasks to run on the command line at runtime, you would have to specify them in the yaml config. Now you can do ./deepy.py evaluate.py -d configs 20B --eval_tasks lambada wikitext hellaswag
• Fixes some minor spelling errors (adaptor vs. adapter)