Introduce reduce operator

[muellerzr]

Introduce a reduce Op to Accelerator

What does this add?

This PR introduces a reduce operator, based on catalyst's implementation + some nuonces for TPU's.

Who is it for?

Users who want to have something akin to .reduce() for calculating their metrics

Why is it needed?

Was requested in #215

What parts of the API does this impact?

User-facing:

Introduces Accelerator.reduce

Internal structure:

Adds a reduce function in operations.py

Basic Usage Example(s):

accelerator = Accelerator()
loss = F.binary_cross_entropy(input, target, reduction="none")
loss = accelerator.reduce(loss, "sum")

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[sgugger]

Thanks for adding this. Should we support nested structures of Tensors instead of just Tensors?