BUG: Fix np.insert for inserting a single item into a structured array

Note that there are some object array special cases because of allowing
multiple inserts. `np.array(..., dtype=object)` is not always clear.

BUG: don't overwrite input percentile arrays

Following inequality causes wrong counting at the edges and can be avoided by
making the edge array of the same type as the input data.

In [1]: np.around(np.float64(6010.36962890625), 5)
Out[1]: 6010.3696300000001
In [2]: np.around(np.float32(6010.36962890625), 5)
Out[2]: 6010.3701

Closes gh-4799

When `x` has more than one element the condlist `[True, False]`
is being made equivalent to `[[True, False]]`, which is correct.
However, when `x` is zero dimensional the expected condlist is
`[[True], [False]]`: this commit addresses the issue. Besides,
the documentation stated that there could be undefined values
but actually these are 0 by default: using `nan` would be desirable,
but for the moment the docs were corrected. Closes #331.

ENH: Speed improvements and deprecations for np.select

The idea for this (and some of the code) originally comes from
Graeme B Bell (gh-3537).
Choose is not as fast and pretty limited, so an iterative
copyto is used instead.

Closes gh-3259, gh-3537, gh-3551, and gh-3254