GeezerVison

Backport of #32545.

PR summary

It's possible to crash ufunc.accumulate by passing an out array that does not match the shape of the input. The fix for that is to check the out shape before relying on it matching.

It's also possible to trigger a read of uninitialized memory while constructing the error for e.g. np.array(b"1").cumsum(dtype="timedelta64[D]", out=np.empty(1)). The fix is to zero-initialize the out_dtypes array.

Also adds new tests for these cases.

AI Disclosure

I triggered crashes related to this during AI-assisted code review. It is not terribly likely for a user to hit these in the real world but I figure it's worth fixing.

I used an AI model to iterate on and review the correctness of the fixes and help write the tests.