v0.9.1: trim yupi cross-check to the random walk; drop the Lévy example

[HeliCorgi]

The bundled Lévy example was removed (kept only the random-walk cross-check). The Lévy radial profile blew Rmax out to ~4920 bins, making tail-shape classification (peak-count O(R^2)) effectively hang, and the verdict was not robust (keep the tail -> slow levy; truncate for speed -> superdiffusive). Lévy / anomalous-diffusion detection is a mature field with dedicated tools and this kernel's tail-shape path offers no advantage, so it is now out of scope. The random-walk check stays: independent Brownian motion -> diffusive / MONOTONE / p1.0, fast and clean.

• Delete examples/yupi_levy2d_radial.npz; no remaining references anywhere.
• generate_yupi_examples.py: drop the Lévy generation/save/reload block; now produces exactly one npz and prints one classification (diffusive).
• tests/test_smoke.py: drop the Lévy LEVY assertion; keep the random-walk file-based test and the pure-NumPy helper unit tests (Gaussian sigma2-linear and the controlled-Cauchy powerlaw one) unchanged.
• kernel_viewer/io/trajectories.py: keep ensemble_to_radial_kernel as a general isotropic utility but reword its docstring -- no longer steers users to it for Lévy, and warns that heavy tails make Rmax large/slow; points to dedicated tools.
• README: cross-library section keeps only the random walk + an honest "Lévy out of scope" note. kernel_viewer core + its output unchanged.
• Version -> 0.9.1.