Handling of zero in log shared axes depends on whether axes are shared

[anntzer]

This is a regression in master compared to 1.5.0.

import numpy as np
import matplotlib.pyplot as plt
ax0 = plt.subplot(221)
ax1 = plt.subplot(222)
ax2 = plt.subplot(224, sharex=ax1)
for ax in [ax0, ax1, ax2]:
    ax.loglog(np.arange(10), np.random.rand(10))
plt.show()

(In 1.5.0 all axes set their xlims ignoring the x=0 point).