logsumexp stabilization can make graph fail with zero-sized arrays

[ricardoV94]

Description

import numpy as np
import pytensor.tensor as pt

x = pt.tensor("x", shape=(0, 2))
pt.logsumexp(x).eval({x: np.zeros((0, 2))})   # ValueError: Input of CAReduce{maximum} has zero-size on axis %d

Interestingly, scipy has the same failure point

import numpy as np
import scipy

scipy.special.logsumexp(np.zeros((0, 2)))  # ValueError: zero-size array to reduction operation maximum which has no identity

This is arguably an edge case. I think a good compromise is to reject the rewrite when the static size is 0, but not otherwise.

[ricardoV94]

Would the problem go away if we treat max([]) == -inf (at least for logsumexp)?