[Bug] Scalar dtype incompatibility with numpy

[kaushikcfd]

>>> a_np = np.random.rand(10, 4).astype(np.float32)
>>> np.minimum(a_np, 255).dtype
dtype('float32')
>>> pt.minimum(pt.make_data_wrapper(a_np), 255).dtype
dtype('float64')

[kaushikcfd]

We should probably use np.min_scalar_type in get_common_dtype_of_ary_or_scalars?

[inducer]

Jax is another important precedent here. They have put considerable thought into their type promotion semantics. It might be reasonable to follow their lead, which I think might, in this case, match the semantics you're after.