Use soft thresholding for an infinite firm upper threshold

[AHMETHAKANBEZIR1]

threshold_firm documents that an infinite upper threshold gives soft-thresholding, but threshold_firm([-3., 0., 3.], 2., np.inf) currently returns [nan, 0., nan] instead of [-1., 0., 1.]. The finite expression computes an inf / inf scale.

Use the existing soft-thresholding implementation for an infinite upper threshold, after the existing threshold validation. No change to the finite firm-threshold formula is needed.

Validation on Python 3.12.14/NumPy 2.5.4/Windows MSVC build:

• Eight new cases cover float32/float64/complex64/complex128, lower thresholds zero and two, zeros, dtype behavior and unchanged inputs. Against the byte-identical main thresholding module in the baseline wheel, all eight fail; four existing tests selected by -k infinite pass.
• Real-image example: the built-in camera image, wavedec2(..., 'db2', level=2), and the first horizontal detail subband of shape (130,130), with lower threshold20 and upper thresholdinf. Main produces 16,900 NaNs; the fixed function produces zero NaNs and exactly matches soft-thresholding. This is a locally constructed image-processing reproducer, not a claimed production incident.
• Rebuilt and installed the fixed wheel. Full compiled package tests: 1,060 passed, 2 skipped.
• Applicable configured prek hooks and git diff --check pass.

Existing open #881 and #883 address equal thresholds and hard-region indexing respectively. Their changes do not cover the documented infinite-upper-threshold path. The baseline numerical module matched main c9b542b; its wheel included the independent version-metadata adjustment from #887. This PR is based directly on main and does not depend on #887.

Other platforms were not tested. Codex prepared and validated this AI-assisted contribution autonomously; no independent human review is claimed.