Fix format_cursor_data for values close to float resolution.

Previously, moving the cursor over
`plt.imshow(1 + np.random.randn(10, 10) * 1e-15)` would result in
`ValueError: math domain error` (when trying to compute
`math.log10(delta)` with `delta = 0`).  This is because the full range
of normalization implied by the image above (~1e-15) is representable
with float precision, but the spacing between individual colors (256x
less) is too small to be represented, and thus rounded to zero.

In general, I consider that such floating point errors should be fixed
on the user side, but this specific error seems worth fixing as the 1)
the image is still rendered correctly, 2) errors in the mouse event
handler (rendering the cursor text) are somewhat obscure for end users,
and 3) it used to work before Matplotlib 3.5.

Author: anntzer

lib/matplotlib/cbook/__init__.py

@@ -2212,6 +2212,10 @@ def _g_sig_digits(value, delta):
     Return the number of significant digits to %g-format *value*, assuming that
     it is known with an error of *delta*.
     """
+    if delta == 0:
+        # delta = 0 may occur when trying to format values over a tiny range;
+        # in that case, replace it by the distance to the closest float.
+        delta = np.spacing(value)
     # If e.g. value = 45.67 and delta = 0.02, then we want to round to 2 digits
     # after the decimal point (floor(log10(0.02)) = -2); 45.67 contributes 2
     # digits before the decimal point (floor(log10(45.67)) + 1 = 2): the total

lib/matplotlib/tests/test_image.py

@@ -341,6 +341,7 @@ def test_cursor_data():
         ([[10001, 10000]], "[10001.000]"),
         ([[.123, .987]], "[0.123]"),
         ([[np.nan, 1, 2]], "[]"),
+        ([[1, 1+1e-15]], "[1.0000000000000000]"),
     ])
 def test_format_cursor_data(data, text):
     from matplotlib.backend_bases import MouseEvent