onnx: ReduceMin/Max Ops

[AnubhabB]

Adds support for ReduceMin and ReduceMax Ops. Found these in use in a Detectron2 exported model.

Implementation
Supports Ops version 18+ Backward compatible as far as I can tell

Open Questions

• Backward compatibility: Op version 13 -> 18 has moved the axes from attributes to input. Though it's trivial to handle this with get_attr_opt::<&[i64]>(node, "axes") fallback, not sure if that's the right way of going about this!
• The Op definition allows for empty set as input and prescribes -INF/ +INF or Datatype::Min/ Datatype::Max for the ops. Reference. I'm not sure if the return is supposed to be a scalar?

@LaurentMazare, any thoughts around these open questions?

[LaurentMazare]

• Backward compatibility: Op version 13 -> 18 has moved the axes from attributes to input. Though it's trivial to handle this with get_attr_opt::<&[i64]>(node, "axes") fallback, not sure if that's the right way of going about this!

Falling back to taking the axis from the input if they are not present in the attributes sounds pretty reasonable to me.

• The Op definition allows for empty set as input and prescribes -INF/ +INF or Datatype::Min/ Datatype::Max for the ops. Reference. I'm not sure if the return is supposed to be a scalar?

I guess we can just error out in this case for the moment, and if anyone runs into this issue then we can have a look at what the failing onnx model actually expects.

[AnubhabB]

Thanks @LaurentMazare.

I've amended the PR with the following:

1. Fetching axes defaults to the current standard (via input) and falls back to attribute for backward compatibility.
2. Errors out on empty Tensors
3. Tests for backward compatibility and empty set error.

Guess this is good to go!

[LaurentMazare]

Thanks!