add faster msa preprocessing

[RaphaelBouvet]

Hi,
Thank you for developing and releasing Tranception :)

During testing of this model, i noticed that the MSA_processing in tranception/utils/msa_utils.py was a limiting step if the msa was too large.

This PR adds a Fast_MSA_processing class with improved speed 🔥 at the cost of more memory.
For example for a msa with 21 k sequences:

    fast processing : 13 sec
    base processing : 472 sec

Instead of doing sequences by sequences comparisons in the original code, i parallelize the calculation.
For big msa, doing the whole comparisons at once is not possible so i used sub_arrays to split the calculation.

• The resulting weights are identical compared to the prereleased weights in my tests.
• The memory usage can be adjusted manually by changing the size of the subarrays (maybe we can adjust depending on user ram)
• The code might not work if there is empty sequences in the msa (not tested)

I am sure there is a better/faster way to do this calculation but this method worked for me

Do not hesitate if you have any questions,
Best wishes,
Raphaël