Numba doesn't respect scalar types

[ricardoV94]

Description

Example revealed in #1406

import numpy as np
import pytensor.scalar as ps

x = ps.float64("x")
y = x.astype("float32")
y.eval({x: np.int64(0.0)}, mode="NUMBA").dtype  # AttributeError: 'float' object has no attribute 'dtype'

[ricardoV94]

As per the devs it actually respects internally, just not when returning to python. We can tweak the tests to use tensor_from_scalar perhaps?