Add tensor_ops::prelu and nn::PReLU

[leofidus]

Lately I've tried porting various computer vision networks to dfdx, and one function that seems quite popular in that field but isn't supported yet is PReLU. For now I just stay away from networks that use it, but it would be nice to have it available in the future.

https://pytorch.org/docs/stable/generated/torch.nn.PReLU.html#torch.nn.PReLU

ParametricReLU is bascially a LeakyReLU where the negative slope is a learned parameter, so there is going to be some overlap with the LeakyReLU implemenation (#287).

[TimerErTim]

Can this really be implemented on a tensor level? LeakyReLU should be no problem but PReLU, which effectively has learnable parameters, is not really possible on purely tensors, is it? nn::PReLU should be easy to do on the other hand.

Edit: Have been thinking about this for some time. Maybe we could make tensor_ops::LeakyReLU use a simple f32 for the slope a whereas tensor_ops::PReLU uses a scalar or (to stay aligned with PyTorch) a vector. This way, a can be backpropagated through in the PReLU operation whereas in LeakyReLU it can't (which is basically it's only major difference).

[opfromthestart]

This should be closed

[coreylowman]

Yep, thanks!