Provide half-away-from-zero rounding mode on Tensor::round

[josevalim]

🚀 Feature

Currently Tensor::round rounds half-to-nearest-even, however, some platforms and tools round half-away-from-zero by default. One of such examples is XLA, which uses half-away-from-zero by default (and half-to-nearest-even is emulated in libraries like JAX).

Operations like Tensor::div already support rounding_mode as an option, so hopefully a similar option - but with different values - can be added to round.

Motivation

I am writing bindings for both PyTorch and XLA for the Elixir programming language and one of the discrepancies we found were the rounding operations. Both XLA and Elixir rounds half-away-from-zero, which is inconsistent to PyTorch default behaviour (which actually aligns with Python 3).

Pitch

The current behaviour rounds both -0.5 and 0.5 to 0. A new rounding option should round them to -1.0 and 1.0 respectively.

Thank you!

cc @yf225 @glaringlee

[jeffreyksmithjr]

Seems like something possibly small enough to be an enhancement, but I'm not the person to say definitively. Added a label to indicate that this is a request specifically to expose this in libtorch, although obviously it would be beneficial to have consistency across the Python and C++ frontends. Just clarifying that @josevalim 's need is in calling into the C++ API.