Implement Multivariate mixtures

[ricardoV94]

Our mixture logprob graphs are tailored for univariate mixtures, by either relying on rv_pull_down rewrites that only work for univariate random variables or assuming there is a 1-1 mapping between the shape value variable and the shape of each random variable component, which is not the case for multivariate distributions.

aeppl/aeppl/mixture.py

Line 335 in 3331081

rv_m = rv_pull_down(rv[m_indices])

aeppl/aeppl/mixture.py

Lines 344 to 347 in 3331081

	logp_val += ifelse(
	at.eq(indices[0], i),
	comp_logp,
	at.zeros_like(value),

aeppl/aeppl/mixture.py

Lines 317 to 320 in 3331081

	original_shape = shape_tuple(comp_rvs[0])
	bcast_indices = expand_indices(indices, original_shape)
	logp_val = at.empty(bcast_indices[0].shape)

The meta information present in RandomVariable.[ndim_supp, ndims_params] and the logic in aesara.tensor.random.utils.broadcast_params`s should give us the tools to infer the right base log-probability shape.

This could also be aided by aesara-devs/aesara#695