Add topk unary op

[WilliamTambellini]

Summary

Returns the k largest elements of the given input tensor along a given dimension.
eg:
https://pytorch.org/docs/stable/generated/torch.topk.html
https://arrayfire.org/docs/group__stat__func__topk.htm#gsc.tab=0

Problem statement

Returns the k largest elements of the given input tensor along a given dimension.

Preferred solution

No preference: single nD input, could compute results into 2 given dest tensors: one for top values, one for top indices.
Ref:
https://github.com/intel/x86-simd-sort/
https://arrayfire.org/docs/group__stat__func__topk.htm#gsc.tab=0

[igorsafo]

Hi @WilliamTambellini ,
oneDNN doesn't try to optimize all operations available in DL frameworks, because it is not scaleble (There are thousands of operations in different frameworks with different semantics across them). Most of operations are good enough already or do not take much time comparing to overall execution time or used in one or two models. This operation is one of these operations and so far we haven't seen any models where it becomes a bottleneck and can be optimized further across multiple frameworks. It also applies to #1697 .

Here is a quote from Contributing.md:

oneDNN focuses on functionality that satisfies all of the following
criteria:

1. Performance: the functionality has material impact on a workload level.
   In other words, this means that for a new primitive it should be
   demonstrated that it brings visible performance improvement to some
   workload.

2. Generality: the functionality is useful in a wide range of deep learning
   applications. This implies that when introducing a new primitive, its API
   needs to be general enough to be integrated into multiple deep learning
   frameworks that have similar functionality.

3. Complexity: it is not trivial to implement the functionality directly in
   a deep learning application.

[igorsafo]

Closing as stale. Feel free to reopen with additional information.