trace fn does not respect random seed for stochastic "Deterministics"

[ricardoV94]

Description

import pymc as pm

with pm.Model() as m:
    x = pm.Bernoulli("x", p=0.5) * 0 # Force it to be zero
    pm.Deterministic("y", x + pm.Normal.dist())

    idata1 = pm.sample(tune=0, draws=1, random_seed=1)
    idata2 = pm.sample(tune=0, draws=1, random_seed=1)
    
print(
    idata1.posterior["y"].sel(draw=0).values,
    idata2.posterior["y"].sel(draw=0).values,
    sep="\n"
)
# [-0.39110838 -0.10602868 -0.64066713  0.59270806]
# [ 1.30907711 -0.1391542   0.52679751  0.47189673]

We could pass the rng to the inner compile_pymc. But it's still a bit silly to compile the same function per chain. We could compile the function once and pass it to the trace. After #7578 this should be allowed. We just need to make sure we copy the function with new shared RNGs.

Alternatively, we could defer to compute_deterministics, and move the deterministics computations out of the sampling loop? This would come at a higher memory cost when users choose not to store certain free variables in the trace.