Use SeedSequence spawning in Model.next_rng

[ricardoV94]

Numpy docs advise using SeedSequence for spawning independent bit_generators: https://numpy.org/doc/stable/reference/random/parallel.html

Their method should be quite more robust to collisions than our naive random.integers(2**30) strategy, we use here and in a couple other places:

pymc/pymc/model.py

Line 1047 in bae5087

new_seed = self.rng_seeder.randint(2**30, dtype=np.int64)

import math

unique_seeds = 2**30
for n_seeds in (10, 100, 1_000, 10_000, 100_000):
    # birthday paradox probability: https://en.wikipedia.org/wiki/Birthday_problem
    p_collision = 1-math.perm(unique_seeds, n_seeds) / unique_seeds**n_seeds
    print(f"{n_seeds=}, {p_collision=}")

n_seeds=10, p_collision=4.1909515080540416e-08
n_seeds=100, p_collision=4.610036260510597e-06
n_seeds=1000, p_collision=0.0004650875835883195
n_seeds=10000, p_collision=0.04549425469529611
n_seeds=100000, p_collision=0.9905023499278603

Similar to aesara-devs/aesara#936

[ricardoV94]

Waiting for #5757 to be able to use Aesara's RandomStream (probably subclassing it so that it doesn't create default_updates).

This will also simplify the construction of SymbolicDistributions so that we can pass a RandomStream directly to them instead of having to ask how many RandomState variables it needs in advance:

pymc/pymc/distributions/distribution.py

Lines 472 to 474 in c18e6ce

	# Instead of passing individual RNG variables we could pass a RandomStream
	# and let the classes create as many RNGs as they need
	rngs = [model.next_rng() for _ in range(cls.num_rngs(*args, **kwargs))]