In numpy.gradient, convert integer array inputs to float64 to avoid
unwanted modular arithmetic.

Closes gh-15207.

Fix wrong multiplier for /proc/meminfo, and do style cleanups.

This reverts commit 3a9a63f81aa37b48bc3eb4efdc2e695210805aa5, reversing
changes made to 7999f7c038a50d79b06edf2d45eb5ff2c25a5a42.

A lot of problems have been exposed by this change, so best to leave it
out of the 1.18.x series and work on it in master.

MAINT: Only copy input array in _replace_nan() if there are nans to replace

Using generators instead of full-blown lists
Using set for search instead of list
Using min to get single element insteaf of sorting full list

DEP: issue deprecation warning when creating ragged array (NEP 34)

This PR allows the axis keyword in expand_dims to be a tuple of ints. Previously, axis could only be an int.

This issue was previously discussed in gh-12290 and the changes are based on gh-12290 (comment).

This PR also removes the deprecation added in v1.13 (2017-05-17), where previously axis could be outside of the range (-a.ndim - 1) <= axis <= a.ndim. Such an axis value will now raise an AxisError. Please let me know if it's too soon to remove this deprecation (I could not find any dev docs stating the length of the numpy deprecation cycle).

Closes gh-12290.

Address gh-14142 for the 1.18 release: warn when saving a dtype with metadata that cannot be loaded.