ML specific operators most needed about matx HOT 1 CLOSED

@drnikolaev can you please describe how you use argmin/argmax in real code? One of the challenges of these functions is we supported batched reductions across an arbitrary number of dimensions, and the indices returned from argmin/argmax must also return the appropriate number of batched indices. For example, if you reduce a 4D tensor to 3D using max, besides the values stored in a 3D tensor, we would need to output another 3D tensor containing 4D coordinates (or 3D compressed/relative offsets). You can see where this would be problematic for these large tensors. Alternatively, we could enable this feature only for smaller reductions, (for example, anything going to 0D would be easy).

from matx.