Make quantized relu more flexible with quant params and use nnlib kernel on HiFi (#4530)

Summary:
Pull Request resolved: https://github.com/pytorch/executorch/pull/4530

As titled. This diff removes the requirement for inputs and outputs of ReLU to share quantization parameters. That should improve the numerics and allow less `requant` nodes in the graph. Since the nnlib kernel does that and is much faster on HiFi, it's a good deal all around.

Reviewed By: hsharma35

Differential Revision: D60696710

fbshipit-source-id: 4fe3faef607b252526cb5aa3a83d064084ba454e