Reject a classifier whose probabilities have a symbolic shape

[LeSingh1]

classify's type inference only checks that the number of class labels matches the size of probabilities when that shape is fully known. With a flexible input shape the check is skipped, so conversion succeeds and returns a model the Core ML runtime cannot compile:

Error: Unable to parse ML Program: in operation classify_1:
Classifier probabilities must have a fully known shape.

This fails for every input shape, including the ones where the symbol resolves to a size that matches the labels, so there is no shape at which the returned model works. A fixed non-matching batch already raises at conversion time; this makes the flexible case fail there too instead of at prediction.

Repro on main (converts, then every predict fails):

@mb.program(input_specs=[mb.TensorSpec(shape=(get_new_symbol(), 4))])
def prog(x):
    return mb.softmax(x=x, axis=1)

ct.convert(
    prog, source="milinternal", convert_to="mlprogram",
    inputs=[ct.TensorType(name="x", shape=ct.EnumeratedShapes([(1, 4), (4, 4)]))],
    classifier_config=ct.ClassifierConfig(["a", "b", "c", "d"]),
    minimum_deployment_target=ct.target.iOS16,
)

Added a parametrized test over enumerated and range-dimension shapes; both fail on main.

Refs #2764