Remove TensorIterator::Builder (#23329)
Summary:
The builder pattern doesn't seem to work well with return-value-optimization.
This saves ~100 ns in the construction of TensorIterator::binary_op.
```
import torch
x = torch.rand(1)
y = torch.rand(1)
z = torch.rand(1)
%timeit torch.add(x, y, out=z) # ~1.76 us vs ~1.88 us on my machine
```
cc resistor zheng-xq
Pull Request resolved: https://github.com/pytorch/pytorch/pull/23329
Differential Revision: D16495070
Pulled By: VitalyFedyunin
fbshipit-source-id: 8ce116075fa4c7149dabfcdfa25885c1187c8e2f