Add Tensor.<func> methods like reduce, activation functions, etc

[albertz]

PyTorch provides all tensor functions also on the Tensor type itself, doc.

This includes:

• all the activation functions (abs, tanh, log_softmax, etc)
• sum (like reduce_sum in TF, or our reduce)
• matmul, etc
• transpose, reshape, etc
• basically everything...

TF does not do this. However, it can often lead to easier code (in my opinion). Examples:

• indices.gather(axis=k_dim, position=i)
• indices.range_in_axis(axis=k_dim)
• x.length(axis=spatial_dim)
• num.reduce(mode="max", axis=num.shape_ordered)
• (a - b).abs().reduce(mode="mean", axis=axis)
• data.length(axis=time_dim).reduce(mode="max", axis=nn.batch_dim)

So first we should clarify if we want this or not.

Then on the technical question how to do it:

I think PyCharm do this automatically on-the-fly when the module is loaded. But I think they provide (auto-generated) pyi files such that autocompletion works with PyCharm etc.

We could also load it on-the-fly. But when we auto-generate pyi files, I think instead we could directly auto-generate some automatic tensor mixin type or so, and then have it more direct.

Or we could explicitly/manually add all the functions. Currently we already do that manually for all the base arithmetic functions like __add__ etc.