Generalize local_sum_prod_alloc

[ricardoV94]

Sum(Alloc(x, n)) (e.g. broadcast_to(...).sum(), or sum(ones((5, 3)) * x)) wasn't simplified to x * n / sum(x) * 5 unless the alloc'd value was a scalar constant. The old local_opt_alloc extracted a constant via get_underlying_scalar_constant_value, which was an inherited Theano restriction rather than a real requirement.

Rewrite changes (pytensor/tensor/rewriting/math.py)

Renamed local_opt_alloc -> local_careduce_of_alloc and generalized it along three axes:

1. Any value, not just scalar constants. The old code only fired when the alloc'd value was a scalar constant (it called get_underlying_scalar_constant_value). The new code has no value precondition at all: the only gate is structural (the reduction's input must be an Alloc). It fires for any value (scalar, symbolic, or holding real data on the kept axes, e.g. a batch dim from vectorization), and uses each reduced axis's broadcastability to decide whether to factor it out as a size multiplier or reduce it on the value.

2. Push the reduction through the alloc. Reduced axes are split into the ones the alloc broadcasts (a brand-new dim, or a size-1 dim the alloc expands) vs. the ones where the value has real data. Broadcast axes are factored out; real axes are reduced on the smaller value via op.clone(axis=...). It never bails except on non-Alloc input, and kept broadcast dims are re-materialized by an output alloc.

3. All seven CAReduce ops, not just Sum/Prod: Sum -> multiply by prod(broadcast sizes); Prod/ProdWithoutZeros -> raise to that power; Max/Min/All/Any (idempotent) -> broadcast axes simply dropped, no factor.

Closes #378

[ricardoV94]

I'll tweak the new tests to be more in line with #2103