Numba `solve_triangular` returns wrong results when `overwrite_b=True`

[ricardoV94]

Description

Revealed in #1226

import numpy as np
import pytensor
import pytensor.tensor as pt
from pytensor.tensor.slinalg import SolveTriangular

rng = np.random.default_rng(1)
a_test = rng.normal(size=(2, 2))
a_test[1, 0] = 0
b_test = rng.normal(size=(2, 2))

op = SolveTriangular(trans=0, unit_diagonal=False, lower=False, check_finite=True, b_ndim=2, overwrite_b=True)

a_pt = pt.matrix("a", shape=a_test.shape)
b_pt = pt.matrix("b", shape=b_test.shape)
out = op(a_pt, b_pt)

py_fn = pytensor.function([a_pt, b_pt], out, accept_inplace=True)
numba_fn = pytensor.function([a_pt, b_pt], out, accept_inplace=True, mode="NUMBA")

(
    py_fn(a_test.copy(), b_test.copy()),
    numba_fn(a_test.copy(), b_test.copy()),
)

# (array([[ 1.64016796,  2.35184226],
#         [ 0.41204025, -0.44593092]]),
#  array([[ 3.4341482 , -0.34253317],
#         [-0.49356511, -0.44593092]]))