correct num_proc logic

[drcege]

Currently, when OP-specific num_proc argument is not set, it will not automatically fall back to global np.

• This is a bug accidentally introduced in a previous commit, now fixed
• Also, fix an edge case where the rank is None in the OCR operator when np=1

[garyzhang99]

In the previous refactor PR #359, we set all op's ray mode to task, which will cause CUDA OOM error.
We need to add back ray actor mode for op that use gpu.