Fast, differentiable fingerprinting dictionary matching

[fzimmermann89]

It would be nice to have a Module that does cosine similiary and dotproduct dictionary matching.
It should take care of batched simulation of the data.
It should use a (GPU) efficient matching, maybe LSH or using faiss.
It should implement two different versions of for the backward:

• running the signal model again at the match to obtain gradients
• using a trick such as softmax/gaussian weighting/... over k nearest neighbours or something similar.

This can then be used for MRF or as initial value estimator for other reconstruction methods.

Ideally, it should keep track how often a dictionary entry was matched and have prune version that removes unused matches.
This would make it really useful for qMRI networks that work with epg signal models

[fzimmermann89]

First would be a simple dictionary module.

On creation, we specify a signal function in the init.
Then, we have a function to append new simulations to the dictionary.
This function takes in as many tensors as the signal function has parameters, simulates the signals, and stores signals and parameters.
The forward would then do the matching by dotproduct and argmax.
The reason for the split of signal function and creation of the dictionary is to allow here for future expandability.

We should try to use this function in the qMRI example notebooks and maybe add a fingerprinting example using a phantom.
For this, I think it would make sense to start the development onto of either the epg or the cmrf branch and once it is working nicely, move to code over to main.

(As a warning: Please be carfule not to commit large files to git, such as raw data, images, etc)