[MRG] Stabilize simplex projection against large common offsets - #880

Summary

Fixes #879. Simplex projection is invariant to a common offset per column. Center floating inputs before sorting and thresholding so large equal float32/float64 inputs retain unit mass rather than producing all zeros. Integer inputs retain their existing path. This does not change sparse-projection handling or validation for nonpositive simplex radii.

Validation

Current head 38faa64 includes master 0b6f8d5. The only merge conflict was RELEASES.md; both entries were retained. The PR production and regression files are unchanged from the previous head.

• Full utility module and doctests: NumPy 124 passed; NumPy/PyTorch 2.9.1 CPU 190 passed; NumPy/TensorFlow 158 passed, 32 existing skips, one median-detachment warning. Each process verifies its actual backends and source checkout; Torch and TensorFlow run separately.
• The regressions use analytic simplex values, ordinary-range controls and the tangent-space gradient reference. Earlier original-function runs reproduced eight failures and twelve controls across NumPy and PyTorch; those are historical baseline checks, not new runs on this update.
• All applicable configured pre-commit hooks for the PR files and diff checks pass.

Windows/Python 3.12.14. Existing native POT extensions were reused; the upstream changes touch no native extension source. No new full-package, local full documentation/gallery, GPU, JAX, CuPy or compilation validation is claimed. Historical focused API documentation passed; current-head upstream checks are pending.

Contribution disclosure

Prepared, investigated and tested autonomously by OpenAI Codex on behalf of AHMETHAKANBEZIR1. No independent human review has been performed.