Feature request: Weighted Square Loss

[marcobonici]

I would like to use a Weighted Mean Square Loss

$$ \mathrm{WMSL}=\frac{1}{N}\sum_{i=1}^N w_i(y_i-\hat{y}_i)^2 $$

This is very close to the standard Mean Square Loss, with the additional Weight array that can be used to give a higher weight to one of the output features.

Something similar can also be done for the absolute square loss, in principle.

After a chat with @chriselrod and @ChrisRackauckas I decided to open an issue and a related PR.
I have forkerd the repo and I am willing to implement this by myself and if you are satisfied merge it.
I should have done most of what is needed, but a bit of guidance would be useful, since I never did something like that before. Basically, right now I adapted the code for the Mean Square error already present in the repo.

Here is the link to my forker version. There are just a few lines of code I added in the loss.jl part. Today I'll check more deeply what I did, but mainly I have a doubt on the target function definition.

[chriselrod]

If you'd like feedback, please open the PR. That makes it easier to see the diff.

I think starting with the MSE code is good; WMSE should be very similar.

[marcobonici]

#140
I created the PR as requested.
I want to stress that there are possibly things that I didn't do in the correct way 😅

Thanks again for your help, I am waiting for your suggestions and feedback!