n3fit prepro initialisation

[comane]

This pull request introduces changes to the n3fit package for better initializer handling.The most important changes include adding a new initializer and modifying the weight generation logic

Initializers update:

• n3fit/src/n3fit/backends/keras_backend/MetaLayer.py: Added RandomNormal to the list of available initializers. ([n3fit/src/n3fit/backends/keras_backend/MetaLayer.pyL11-R19](https://github.com/NNPDF/nnpdf/pull/2249/files#diff-b1e9997427e416cda90e812413c31a8c9c0748eafea0276fc297c9a1e0f6e04eL11-R19))

Weight generation logic improvement:

• n3fit/src/n3fit/layers/preprocessing.py: Updated the generate_weight method to use a dictionary for initializer parameters and apply constraints only if the initializer is random_uniform. ([n3fit/src/n3fit/layers/preprocessing.pyL79-R87](https://github.com/NNPDF/nnpdf/pull/2249/files#diff-29a4fb3f4aba49242bccaf2c84b6c891a276e329b2ef1d7b8c20575ce83e0f88L79-R87))

Note

Based on the introduced changes the flav_info dict in the runcard would change to something like:

flav_info: [
    {
        "fl": "sng",
        "trainable": False,
        "smallx": {ini_name: "random_normal", mean: 1, stddev: 0.001}, #minval=1.099, maxval=1.109}, #[1.099, 1.109],
        "largex": {ini_name: "random_uniform", minval: 2.197, maxval: 2.397}, #[2.197, 2.397],
    },

....

This would however allow for the use of different initialisations

[RoyStegeman]

Somehow I think this has been done, or at least thought of before, but I can't find any results or recall why it's not implemented. What is the reason you want to try this?

Using dict logic if the names of parameters may differ is fine for me.