[Fix #99] Convert wrscatter directions from degrees to radians

[dylanpulver]

wrscatter negates the direction while it is still in degrees and then adds np.radians(90), so the value handed to the polar axes is not an angle in radians at all:

direction = -np.array(direction) + np.radians(90)

The transform the rest of the library uses is theta = radians(90 - direction) - that is what the angles array in WindroseAxes._init_plot computes, and it is what bar, box, contour and contourf all plot against. Measured against that convention, on main:

compass   expected theta   plotted theta   error
      0           1.5708          1.5708     0.0 deg
     45           0.7854          0.5531   -13.3 deg
     90           0.0000          5.8186   -26.6 deg
    180           4.7124          3.7832   -53.2 deg
    270           3.1416          1.7478   -79.9 deg

Only 0 degrees is right, by coincidence. This is the same symptom as #99 ("all the polar plots have 0 angle at North, but wrscatter is drawn like a usual polar plot"): the 2018 fix for that issue added -wd + pi/2, which is the correct expression for wd in radians, but wrscatter is fed degrees. So this finishes that fix rather than changing the intended convention.

Also affects plot_windrose(..., kind="scatter"), since D_KIND_PLOT["scatter"] is wrscatter.

Three image baselines encoded the old placement and are regenerated: output/func/test_wrscatter.png, output/df/test_scatter.png, output/df/test_windrose_np_plot_and_pd_plot.png. Before you trust them: regenerating test_wrbar.png from unmodified main here reproduces the committed file with RMS 0.0000, so this machine's rendering matches whatever produced the current baselines, while the committed test_wrscatter.png was already RMS 5.76 away from what unmodified main draws here, against a tolerance of 10. After this PR that figure regenerates exactly.

Added test_wrscatter_direction_convention, which asserts the plotted theta against radians(90 - direction) rather than against a picture. Ran pytest -s -rxs -Werror tests/ --mpl on Python 3.14 / matplotlib 3.11.1 / macOS: 34 passed before, 35 passed after. Reverting only windrose/windrose.py fails the new test; so does writing radians(direction) without the sign flip. Writing it as -radians(direction) + pi/2 is algebraically the same and also passes - I used the radians(90 - d) form because it reads as the library's own convention.

Not touched: WindroseAxes.scatter itself, which tests/test_windrose_np_mpl_oo.py calls with raw degrees and no transform — a separate question, and changing it would break anyone already passing radians.

AI disclosure: drafted with Claude Opus 5 (claude-opus-5); the numbers above are from runs on this machine.

[ocefpaf]

I'm not super familiar with the code, but consistency with the other plotting methods is definitely a plus! We would need to mint a new major release though b/c, in a way, it is a breaking change.

I'll leave this here to see if we can get more comments from others. If not I plan to take a look using the best of my knowledge of the code and merge.