Deprecate torch.add(tensor, value, other)

[zou3519]

🚀 Feature

It's confusing that add has the scale argument. Could we create a new torch.axpy function that does this?

Motivation

Numpy also makes this separation: they have numpy.add, and scipy.blas.axpy.
This could also make things nicer: add can be treated as an example of a binary pointwise operation like mul, sub, div and not require any special casing for its value argument.

cc @ezyang @gchanan @mruberry @rgommers @heitorschueroff @cpuhrsch

[vadimkantorov]

@zou3519 Empirically, torch.add(torch.rand(3, 3), torch.rand(3, 3)) works, so it seems it does the dispatch correctly if it's used as a binary op. Strange.

[zou3519]

Another reason for deprecating it is that the overload cannot be written nicely with a single python function (without parsing *args and **kwargs).

The signature looks something like add(tensor, value=1, other): the value is optional.

[ezyang]

@colesbury says: it would be good to get rid of all of the weird ambiguous positional arguments, in general. (These cases may all be in deprecated.yaml).

@gchanan says: For example, sub, and all of the fused add ones, like addcmul. So we want a more general solution than just calling this axpy