BUG: PcolorImage handles non-contiguous arrays, provides data readout

[efiring]

Closes #6905, and fixes a bug in the underlying _image.pcolor2 function. It requires that all three input arrays be contiguous, so I appended _contiguous to the converters for x and y.

To verify both the data cursor and the ability to handle non-contiguous inputs,

import numpy as np
import matplotlib.pyplot as plt

x = np.arange(4) ** 1.5
y = np.arange(3) ** 1.5

# comment out for contiguous inputs
x = x[::-1]
y = y[::-1]

z = x[:-1] - y[:-1, np.newaxis]

fig, ax = plt.subplots()
pci = ax.pcolorfast(x, y, z)
print(type(pci))
print(pci._A.shape)
print(pci._A)
print(pci._Ax)
print(pci._Ay)

plt.show()

[efiring]

Lots of unrelated failures: ERROR: matplotlib.tests.test_mathtext.test_mathtext_exceptions

[jenshnielsen]

I fixed the unrelated error on master so cycling this to rerun on current master

[efiring]

@mdboom, on second thought, what I characterized as a bug in _image.pcolor2 is better described as a limitation or requirement. Originally, PcolorImage was ensuring that x and y are contiguous. I broke that by indexing with [::-1] when necessary to make them increasing so I could use np.searchsorted on them. I could have added calls to np.ascontiguousarray, but judged that it was better to let _image.pcolor2 handle this. If you disagree, I will move the sanitization back out to PcolorImage.

[tacaswell]

Do we enforce that x/y are monotonic someplace else?

If we bump the minimum numpy to 1.7, searchsorted takes in a sorter kwarg which is an array with the indexes in sorted order (https://docs.scipy.org/doc/numpy/reference/generated/numpy.searchsorted.html)

[efiring]

No, we have never enforced monotonic x or y for any of the pcolor-like functions; that has implicitly been left to the user. If the user supplies non-monotonic x and/or y the plotted results will be odd, and with this PR as-is the cursor readout will sometimes be hard to predict--but the plot won't make sense anyway, so I don't see that it matters. So it seems to me that you are raising a question that is independent of this PR--should we add more input validation for this category of plot?

[tacaswell]

I am 👍 on this modulo my concern about x/y being monotonic.

[tacaswell]

The travis mac infrastructure is down, merging as it previously passed an the new changes are docstring only.

[tacaswell]

backported to v2.x as fab69a1