Keep is_self_adjoint hint in LinearOperatorScaledIdentity.inverse() - #129948

LinearOperatorScaledIdentity._linop_inverse always sets is_self_adjoint=True. That holds for a real multiplier, but not for a complex one, and adjoint() returns self when the hint is True, so the adjoint of the inverse comes out wrong:

import tensorflow as tf
op = tf.linalg.LinearOperatorScaledIdentity(num_rows=2, multiplier=1j)
inv = op.inverse()                     # -1j * I
inv.is_self_adjoint                    # True
inv.adjoint().to_dense()               # [[-1j, 0], [0, -1j]]
tf.linalg.adjoint(inv.to_dense())      # [[ 1j, 0], [0,  1j]]

The same wrong self is used by inv.H, inv.matmul(other_operator, adjoint=True) and inv.solve(other_operator, adjoint=True), and inv.assert_self_adjoint() fails.

The fix passes self.is_self_adjoint, like _linop_adjoint and the other operators' _linop_inverse already do. For a real multiplier nothing changes: the constructor sets is_self_adjoint=True on its own.

test_scaled_identity_inverse_complex_multiplier compares inverse().adjoint() with the dense adjoint for multiplier 1j, and checks that a real multiplier still gives a self-adjoint inverse. It fails before the fix (graph and eager) and passes after. The whole linear_operator_identity_test.py (2689 tests), linear_operator_diag_test.py and linear_operator_addition_test.py pass with the change. I ran them against tf-nightly 2.23.0.dev20261103 with the patched module, not a source build.