added cosine matching loss

[ankitpatnala]

Description

Added cosine matching loss which is used for constraining cosine similarity between consecutive forecasting latents

Issue Number

Link the Issue number this change addresses:
Closes #2481

[ankitpatnala]

@SavvasMel

[clessig]

Thanks for the implementation. Can we generalise the PR slightly to have a LossLatent (I think it's a regularisation term since it does not involve targets but this is a minor point), with the cosine being a specific implementation (RMSE or other losses are possible as well). We will also need to best structure the part in the model but let's first to this restructuring.

@MatKbauer : (I don't think so since this uses that target_aux_calculator but double-checking)

@sophie-xhonneux : for me this is more a regularisation term since no target are involved. Any opinion on this?

[ankitpatnala]

Thanks for the implementation. Can we generalise the PR slightly to have a LossLatent (I think it's a regularisation term since it does not involve targets but this is a minor point), with the cosine being a specific implementation (RMSE or other losses are possible as well). We will also need to best structure the part in the model but let's first to this restructuring.

@MatKbauer : (I don't think so since this uses that target_aux_calculator but double-checking)

@sophie-xhonneux : for me this is more a regularisation term since no target are involved. Any opinion on this?

Thanks for the comments! I have implemented few changes you proposed. Now I have removed targets and changed the naming to LatentLoss.
With regards to restructuring, I depopulated the forward function by moving the cosine calculation operation to loss/utils.py