Preserve empty tensor shapes in unfolding - #710

Fixes #709.

Unfolding an empty tensor can fail when a retained dimension is zero: reshaping (0, 3, 4) to (0, -1) cannot infer the second dimension, although its original shape determines (0, 12). Empty skipped batch/channel dimensions cause the same problem in partial unfolding and vectorization.

Compute the collapsed dimension from the original, unskipped axes when a known static zero makes inference ambiguous. Keep the existing inference path for other inputs, including unknown dimensions, and preserve axis ordering. The partial vectorization alias inherits the repair; folding and full vectorization are unchanged.

Validation on Windows, Python 3.12.14 / NumPy 2.3.5 / SciPy 1.16.3:

• Full NumPy suite with coverage: 1035 passed, 12 skipped, 1 existing non-strict XPASS, with 35 warnings from original tests. This includes all 34 added regression/control cases.
• The same suite on the unmodified base produces 18 failures, all in new cases; all original test outcomes are unchanged.
• PyTorch 2.8.0 CPU: the affected core module and independent empty/positive shape and value controls pass. These overlap the NumPy cases and are not additional full-suite tests.
• Black 26.10.0 checks both changed files. A separate frozen-source review found no must-fix issues.

The unknown-dimension tests check shape dispatch; they do not execute a TensorFlow graph. Other backend/version/OS/device matrices, documentation and package builds were not run locally.