Support for non-funsor arguments to ops

[fritzo]

This issue tracks changes needed to support non-funsor args as in @ordabayevy's #482, e.g.

ops.sum(Tensor(data), -1, keepdims=True)
# -op-dispatch-> Unary(ops.SumOp(-1, True), Tensor(data))
# -eager-> Tensor(ops.SumOp(-1,True)(data))  # TODO fix batch dimension handling

Plan

Note during the first two refactorings, we may want to try to decouple dispatch tables from the Op hierarchy (but that would be a bigger refactoring, possibly beyond the scope of this issue).

• Refactor ops to allow non-funsor parameters #491 Refactor Op so that the dispatch table is in Op subclasses, not in instances.
• Support ops.sum(data, dim=None, keepdims=False) #490 Demonstrate parametrized ops via ops.sum(data, dim=None, keepdims=False)
• Add ReductionOp, ops.mean, ops.std, ops.var #482 Fix batch dimension handling in Tensor.eager_unary()
• Add ReductionOp, ops.mean, ops.std, ops.var #482 add more ops, e.g. a richer ops.sum, ops.mean, etc.

[eb8680]

Can this be closed now that #482 is merged?

[fritzo]

Yes, thanks for your help @ordabayevy !