[Integrate] Bump LLVM version to 72efe15c2bf262607420f81ab3b3b52f82698b4b (#54922)

- Previously it was assumed that nb::gil_scoped_release release; could
not be used for compiling torch models, due to handling
`PyDenseResourceElementsAttribute` in constant folding. However, there
is no guarantee
that the thread releasing the PyDenseResource will be holding the GIL
(this was addressed in llvm/llvm-project#124832). This PR avoids GIL
deadlocks by releasing the GIL in both code paths.
- Unblocks Python 3.12 support.
- Update lldb20.0.0git to lldb21.0.0git
- Add `funcRange.front().GetBaseAddress()` to the `Function` constructor
call.
- Additional dependencies from LLVM/MLIR are necessary since they are no
longer included transitively.
- llvm/llvm-project#123488 causes new
dependencies on UBDialect.

MAX_GRAPH_API_ORIG_REV_ID: cab252a7205f18c0a1320bc4e0f1689f2731932a

Author: naibaf7

tests/max/graph/test_graph.py

@@ -210,7 +210,7 @@ def test_invalid_operand() -> None:
         ],
     ) as graph:
         input_tensor = graph.inputs[0]
-        with pytest.raises(AssertionError):
+        with pytest.raises(ValueError):
             Graph.current._add_op(rmo.add, [2, 5], input_tensor)
         graph.output(input_tensor)