Add few ops as methods of tensors

[akhilg-nv]

Certain operations (specifically basic shape manipulations - reshape, transpose, permute, etc.) are commonly used as methods of tensors in other frameworks (Torch, Numpy, Jax, etc.) and often "chained" together, e.g. tensor2 = tensor1.method1(args1).method2(args2).method3(args3). Adding these as methods of tensors improves code readability significantly when used in this way, as we've seen when translating demos from Torch to Tripy (the current approach requires tensor2 = tp.method3(tp.method2(tp.method1(tensor1, args1), args2), args3) to write on one line).

We should add such methods as methods of tensors, however, we only want to do this for a few methods where it "makes sense" and not create too many ways to do the same thing for every operation. Some approaches to evaluate this:

• Check what NumPy / JAX support as methods of tensors
• Check demos and see what patterns are commonly using a "chain" of methods on a tensor (i.e. are commonly paired with other related funcs)
• Gut check if it "makes sense" - would you think to call tensor1.ReLU() or tp.ReLU(tensor1)? Assume users read left to right.