refactor: default _op_preserves_axis_identity's fallback to unsafe

After the previous commit's restructuring, this function is only ever
called for ops confirmed to be in _AXIS_REORDERING_ATEN_OPS (transpose,
t, permute) - and the three branches above already cover every member
of that set exhaustively, so the trailing fallback is unreachable in
current operation.

Flip it from True to False anyway, so a future op added to
_AXIS_REORDERING_ATEN_OPS without a matching branch here fails safe
(assumed to move the axis) rather than fails open (assumed to leave it
untouched). Before the prior commit, this fallback was still reachable
by pooling ops falling through with no explicit branch, so it had to
stay True; now that pooling is dispatched to a separate branch in the
caller and never reaches this function, the flip is a no-op in
practice (confirmed: no test outcome changes) but closes the gap for
whatever's added here next.

Author: guru-desh

src/coreai_opt/quantization/_graph/_utils.py

@@ -314,8 +314,10 @@ def _op_preserves_axis_identity(op_node: Node, axis: int) -> bool:
     ``axis`` to a different position between its input and output.
 
     Only called for ops in ``_AXIS_REORDERING_ATEN_OPS`` (transpose, t,
-    permute); the three branches below cover exactly those ops, so the
-    trailing fallback is unreachable in practice.
+    permute), which the three branches below exhaustively cover. The
+    trailing fallback defaults to False rather than True, consistent with
+    this module's fail-safe default: a reordering op this function doesn't
+    yet know how to analyze should never be assumed identity-preserving.
     """
     if op_node.target == torch.ops.aten.transpose.int:
         _, dim0, dim1 = op_node.args
@@ -331,7 +333,7 @@ def _op_preserves_axis_identity(op_node: Node, axis: int) -> bool:
     if op_node.target == torch.ops.aten.permute.default:
         _, dims = op_node.args
         return dims[axis] == axis
-    return True
+    return False
 
 
 def _force_fake_quant_to_per_tensor(fake_quant: FakeQuantizeImplBase, op_node: Node) -> None: