Replace getattr with constant when buffer is a graph output

[john-rocky]

Summary

• remove_getattr_nodes raised when a model directly returned a registered buffer / parameter (e.g. forward(self): return self.buf):

  RuntimeError: my_constant_output2 should not be in the graph outputs.

• The pre-existing flatten_graph_output_values pass already lifts buffer-returning getattr nodes into the graph outputs, so the bare raise left no escape hatch — there is no other op that lowers them.

• Replace each output-producing getattr with a constant node carrying the buffer value (looked up in graph.params). The downstream constant op handler then materializes it as a const var. Intermediate getattr nodes are still dropped as before, since their consumers read directly from graph.params.

Test plan

• New unit tests TestTorchPasses::test_remove_getattr_nodes_with_output_buffer and ::test_remove_getattr_nodes_drops_intermediate in test_passes.py.
• All existing TestTorchPasses tests still pass (7 passed, 1 xfailed).
• End-to-end conversion of the issue repro succeeds and mlmodel.predict({}) returns the original buffer values (verified locally with minimum_deployment_target=ct.target.iOS18, since empty-input models also need PR Raise a clear error when a no-input model targets below iOS18 #2684's spec-version bump).

The original issue author @tritolol proposed essentially this fix in the issue description; this PR implements it with a defensive lookup against graph.params, params propagation through nested blocks, and a unit test.

Fixes #2538.

[TobyRoseman]

@john-rocky - Is this pull request still relevant? The associated issue has been closed. If it is still relevant, please fix the merge conflict.

[john-rocky]

Closing this — #2538 was fixed by #2541 (merged), which makes the same remove_getattr_nodes change. Thanks @tritolol! No longer needed.