Fix crashes on modern NumPy/Matplotlib and low logging-rate logs

[deliphop]

Summary

The project pins numpy 1.11/scipy 1.0/matplotlib 2.0 (2017-2018 era), which have no wheels for current Python, so this fixes compatibility with recent releases instead:

• np.histogram/histogram2d: normed= was removed, use density=
• hist2d(): bin count must be an int, not a float from Python 3 division
• pcolormesh(): explicit shading= is now required; one call mixed bin edges with sample centers across axes, which newer Matplotlib rejects outright (fixed by aligning the smaller axis to centers)
• plt.cm.get_cmap() was removed, use plt.colormaps[...]
• get_shared_x_axes().join(...) was removed, use ax.sharex(...)

Also fixes two crashes specific to logs recorded at a low effective rate (e.g. FCs logging to internal flash with blackbox_rate 1/4, rather than a full-rate SD card log):

• winstacker(): the window count formula can overrun the data array by a few samples on the last window when flen isn't a clean multiple of shift*superpos, producing a ragged array and crashing np.array(). Now clamps the window count to what actually fits.
• plot_all_noise(): sparse analysis windows can leave NaN in the noise spectra, which crashes the LogNorm colorbar. NaNs are now sanitized to 0 before computing color limits, which correctly falls through to the existing "no data" default range.

Test plan

• Ran against the bundled good_tune.BBL / stock_tune.BFL example logs — PID response and noise plots generate correctly
• Ran against a real-world log recorded at blackbox_rate 1/4 on internal flash (126 Hz effective rate, 75% frame decimation) — previously crashed on both fixed issues, now completes and produces sane plots
• Verified with numpy 2.x / scipy 1.x / matplotlib 3.x on Python 3.14