currently indices_obj and indices are the same, but it can change in the futur.

Shape tensor basic2

Consider numpy.memmap a valid type for TensorType

Remove old init parameter that was removed in the past. Recently, we added new one. So there was a conflict.

Also assert that the output dtype if float64.

This is not important as the Mean op is not used by default.

A reshape fix this problem, but return the wrong results.

add infer_shape to AdvancedIncSubtensor and add tests; repair infer_shape of Reshape allowing for constant desired shape with possibly one entry equal to -1 and add tests

Join: added test on second axis;  TensorDot: added axes ((0, 1), (1, 0)); MaxAndArgmax: removed TODO comment