[Feature] empty_like & co. to support the shape and order arguments

[madsbk]

... to match numpy.empty_like(), numpy.ones_like(), etc.

shape should be straightforward to support.
order might be more difficult but at least NotImplementedError could be raised, if the order is being changed.

[manopapad]

I agree, shape should be straightforward to implement.

order currently does not affect the layout of actual instances in cuNumeric. Instead we lay them out at our discretion, according to the requirements of the library that we use (e.g. cuSolver requires column-major, so that's what we do). We can do that because these instances are not immediately accessible to the user code (they can only interact with them through the legate library). So we've pretty much been ignoring the order parameter.

What we should be doing is recording this information, so that if the user code ever asks for the whole array to be collected in one place (which happens implicitly if they try to print it, or iterate over its elements), then at that point we respect the user's wishes. This would also require that when "inline mapping" a region (that's the Legion mechanism we use to make that "local full-data collection" happen) we propagate this order requirement to the inline mapping operation.