Add a baseline CFIS dataset

[EiffL]

This PR adds a first pass of a tf dataset of CFIS-like images, this aims at solving #39.

This is a first pass because we will probably want to add additional variants on that dataset in the future, but it is a starting point

[EiffL]

Super easy to use:

import autometacal
import tensorflow_dataset as tfds
import tensorflow as tf

# We create a function to add noise on the fly for data augmentation
@tf.function
def add_noise(example):
    im_noise = example['obs'] + example['noise_std'] * tf.random.normal([51,51])
    return im_noise, example

dset = tfds.load('CFIS/parametric_shear_1k', split='train')
dset = dset.map(add_noise)

for im, example in dset:
  # im contains the noisy observation
  # example contains psf image, observation without noise, and galaxy magnitude
  ....

[aguinot]

I checked the notebook, I am a bit confused by the results:

• Do you know why autometacal is "slower"? (from the tqdm print)
• The response are different from ngmix and autocal (and they are not negligible).
• Autocal is 1 magnitude better than ngmix, do you think this is because of autodiff? If so, this is very cool!
• I think I would make a test at very high SNR (like 50000-100000) just to make sure everything is working well. Maybe you have already tried?

[EiffL]

Dont look too closely at the notebook. The response computed by autometacal is wrong, cf #26. The tqdm looks slower but that's because I do like 10000 instead of 1000 galaxies.

The main point of this PR is to add the dataset, the notebook is to illustrate how to load and use the dataset, but it's optional, if you think it's not useful I can drop it for now.