BUG: preserve coefficients during inverse multiresolution analysis

[antonsoo]

imra2 and imran add details into the supplied approximation array. Reconstructing twice from the same decomposition therefore adds the details twice, and later filtered reconstructions use a changed approximation:

import numpy as np
import pywt

x = pywt.data.camera()[192:224, 192:224].astype(float)
coeffs = pywt.mra2(x, 'db2', level=2)
approximation = coeffs[0].copy()
first = pywt.imra2(coeffs).copy()
second = pywt.imra2(coeffs)
print(np.max(np.abs(first - second)))  # about 44 before; 0 after
np.testing.assert_array_equal(coeffs[0], approximation)

Copy the approximation before summing in both inverse functions. This retains the existing summation order and allocates one output array. It also allows read-only coefficient arrays.

Validation covers actual camera/RGB-face/ECG data with DWT/SWT, both APIs, db2/sym4 and real/complex single/double precision. All 80 configurations change their input before; none do afterward. All 80 first-reconstruction outputs remain byte-identical, repeated outputs now agree exactly, and all read-only cases complete. An independent sum of attenuated components agrees within 4.32e-5 in float32.

All 32 new regressions fail on main and pass with the fix. The full package suite passes (1084 tests, two skips), as do the project pre-commit hooks. The released 1.10.0 wheel and #847's head reproduce the same mutation; #847 changes annotations only in this module.