Refactor _check_start_shape to evaluate shapes without needing to draw samples

[michaelosthege]

The current implementation catches sampling errors, but in theory we should be able to evaluate shapes without needing random sampling implementations.

Also see #5027

[michaelosthege]

Here's something to start with:

for rvname, dims in pmodel.RV_dims.items():
    rv = pmodel[rvname]
    actual = tuple(rv.shape.eval())
    expected = tuple(
        int(pmodel.dim_lengths[dname].eval())
        for dname in dims
    )
    if actual != expected:
        raise pymc3.ShapeError(f"Something is wrong with RV '{rv}'.", actual, expected)