The purpose of this PR is to continue several threads of discussion
regarding `top_k`.

This follows roughly the specifications of `top_k` in
data-apis/array-api#722, with slight modifications to the API:

```py
def topk(
    x: array,
    k: int,
    /,
    axis: Optional[int] = None,
    *,
    largest: bool = True,
) -> Tuple[array, array]:
    ...
```

Modifications:
- `mode: Literal["largest", "smallest"]` is replaced with
`largest: bool`
- `axis` is no longer a kw-only arg. This makes `torch.topk`
slightly more compatible.

The tests implemented here follows the proposed `top_k`
implementation at numpy/numpy#26666.