Create masks from regex

[rlouf]

This PR adds a function that creates a mask from strings that represent a regex. I also add a aliases for a few frequent use cases:

• floats
• integers
• set of characters

Masks are returned as boolean np.ndarray, and can thus be combined using | (logical or) and & (logical and):

>>> import numpy as np
>>> a = np.array([True, False, True])
>>> b = np.array([True, True, False])
>>> a & b
array([ True, False, False])
>>> a | b
array([True, True, True])

I originally used torch.tensors but decided to use numpy.ndarrays in the end. While this could allow supporting models written with different frameworks, I am still wondering if the conversion to and from the frameworks would slow down the generation due to copying the content of the arrays.

In the Numpy-> Framework direction, jax.numpy.asarray does not copy the input is a numpy.ndarray, and neither does torch.from_numpy. So we should be able to return NumPy arrays here without adding any overhead.

[rlouf]

I benchmarked create_int_mask, create_float_mask and create_char_set_mask against their pure python (no regex) equivalent. The regex version is systematically faster.