Fix broken scatter/fill_between in log-scale axis

I did try to fix this myself, but failed. So I asked Codex (gpt5-high)
and it found the issue and proposed this fix, which I verified works,
and I streamlined its code a bit.

Here's what it had to say:

> - `mplexporter/mplexporter/exporter.py:6-275` now imports `numpy/mpath` and routes collections through `_prepare_collection_data`, which preserves data-space vertices/offsets whenever a non-affine `ax.transData` branch is
>    involved (log scales, semilog, etc.). As a result, `fill_between`, `scatter`, and other `PathCollections` on log axes export with `pathcoordinates/offsetcoordinates == 'data'`, so the rendered polygons obey zooming/panning.
> - `mpld3/tests/test_elements.py:76-103` adds regression tests for both `fill_between` and `scatter` on logarithmic axes to guard against future regressions in the exported coordinate system.

Author: lucasb-eyer

mplexporter/exporter.py

@@ -6,10 +6,11 @@
 """
 import warnings
 import io
+import numpy as np
 from . import utils
 
 import matplotlib
-from matplotlib import transforms, collections
+from matplotlib import transforms, collections, path as mpath
 from matplotlib.backends.backend_agg import FigureCanvasAgg
 
 class Exporter(object):
@@ -235,7 +236,7 @@ def draw_collection(self, ax, collection,
                         force_offsettrans=None):
         """Process a matplotlib collection and call renderer.draw_collection"""
         (transform, transOffset,
-         offsets, paths) = collection._prepare_points()
+         offsets, paths) = _collections_prepare_points(collection, ax)
 
         offset_coords, offsets = self.process_transform(
             transOffset, ax, offsets, force_trans=force_offsettrans)
@@ -279,3 +280,52 @@ def draw_image(self, ax, image):
                                  style={"alpha": image.get_alpha(),
                                         "zorder": image.get_zorder()},
                                  mplobj=image)
+
+
+def _collections_prepare_points(self, ax):
+    # Code is from mpl.collections._prepare_points, but we preserves data-space
+    # vertices/offsets whenever a non-affine ax.transData branch is involved
+    # (log scales, semilog, etc.) ; that's the `contains_branch` test.
+    # As a result, fill_between, scatter, and other PathCollections on log axes
+    # export with pathcoordinates/offsetcoordinates == 'data', so the rendered
+    # polygons obey zooming/panning.
+    #
+    # The first arg is called self, to keep code close to original.
+    transform = self.get_transform()
+    transOffset = self.get_offset_transform()
+    offsets = self.get_offsets()
+    paths = self.get_paths()
+
+    if self.have_units():
+        paths = []
+        for path in self.get_paths():
+            vertices = path.vertices
+            xs, ys = vertices[:, 0], vertices[:, 1]
+            xs = self.convert_xunits(xs)
+            ys = self.convert_yunits(ys)
+            paths.append(mpath.Path(np.column_stack([xs, ys]), path.codes))
+
+        if offsets.size:
+            xs = self.convert_xunits(offsets[:, 0])
+            ys = self.convert_yunits(offsets[:, 1])
+            offsets = np.column_stack([xs, ys])
+
+    def contains_branch(tr):
+        return ax is not None and tr.contains_branch(ax.transData)
+
+    if not transform.is_affine and not contains_branch(transform):
+        paths = [transform.transform_path_non_affine(path)
+                    for path in paths]
+        transform = transform.get_affine()
+
+    if not transOffset.is_affine and not contains_branch(transOffset):
+        offsets = transOffset.transform_non_affine(offsets)
+        # This might have changed an ndarray into a masked array.
+        transOffset = transOffset.get_affine()
+
+    if isinstance(offsets, np.ma.MaskedArray):
+        offsets = offsets.filled(np.nan)
+        # Changing from a masked array to nan-filled ndarray
+        # is probably most efficient at this point.
+
+    return transform, transOffset, offsets, paths