combine should be parallelizable in many cases

Relevant discussion: https://lists.apache.org/thread.html/r9e7d9527eb1d4c9c097c91c010a25dabf4a5f8053d50dc3b6d90d36a%40%3Cdev.beam.apache.org%3E

Currently we require Singleton partitioning for combine() because func might operate on the full dataset, but in many cases func is actually an elementwise method. We should detect this when possible (e.g. when func is an np.ufunc), and/or provide a flag to let the user indicate the function is elementwise.