Fix torch.randint generating values equal to the upper bound (#2568)

[john-rocky]

Summary

torch.randint(low, high, ...) draws integers in [low, high) — the upper bound is exclusive. The lowering used mb.random_uniform (whose bounds are inclusive) and cast the result to int, so the converted model could emit exactly high (issue #2568).

Clamp the cast result to high - 1 with a minimum op so the converted graph matches torch's exclusive upper bound:

-    rand_int = mb.cast(x=rand_uniform, dtype="int32", name=node.name)
+    rand_int_unclipped = mb.cast(x=rand_uniform, dtype="int32")
+    high_int = mb.cast(x=high, dtype="int32")
+    high_minus_one = mb.sub(x=high_int, y=1)
+    rand_int = mb.minimum(x=rand_int_unclipped, y=high_minus_one, name=node.name)

Fixes #2568.

Test plan

New TestRandint::test_upper_bound_is_exclusive (TorchScript + TorchExport frontends; ExecuTorch is skipped because aten.randint.low is not Aten Canonical) asserts the lowered graph wires a minimum clamp into the randint path, so the converted graph cannot produce values >= high.

Verified locally on macOS that the clamp is present in the lowered graph (convert_to="milinternal"):

[torch.export] random_uniform: True | minimum: True

The full test converts to a serialized mlprogram, so it runs end-to-end in CI.

[TobyRoseman]

Is this statement necessary? If so, why?

[TobyRoseman]

I don't understand what the rest of this test is doing. Why not just get the predictions from the converted model and make sure the max element is 100?

[john-rocky]

Done — replaced the MIL op-type walk with a prediction check.

• On the inputs line: it's needed because the TorchScript frontend requires explicit input types, while exported programs already carry them (same pattern as test_tuple_input just above). Added a short comment to that effect.
• On checking the output: I assert prediction.max() < 100. Since randint's upper bound is exclusive, a correct model tops out at 99; before the clamp, a 100k-element draw hits exactly 100 — so it's < 100 rather than == 100.

Verified locally: passes on this branch, and fails on main (the unclamped cast emits 100), so it catches the regression.

[TobyRoseman]

CI: https://gitlab.com/coremltools1/coremltools/-/pipelines/2553884155

If this passes, I will merge.