arrow

ARROW-1156: [C++/Python] Expand casting API, add UnaryKernel callable. Use Cast in appropriate places when converting from pandas

cc @cloud-fan

With this patch we now try to cast to indicated type on ingest of objects from pandas:

```
In [3]: arr = np.array([None] * 5)

In [4]: pa.Array.from_pandas(arr)
Out[4]:
<pyarrow.lib.NullArray object at 0x7f6cf1485d18>
[
  NA,
  NA,
  NA,
  NA,
  NA
]

In [5]: pa.Array.from_pandas(arr, type=pa.int32())
Out[5]:
<pyarrow.lib.Int32Array object at 0x7f6cf1485d68>
[
  NA,
  NA,
  NA,
  NA,
  NA
]
```

I also added zero-copy casts from integers of the right size to each of the date and time types.

Includes refactoring for ARROW-1481.

Author: Wes McKinney <wes.mckinney@twosigma.com>

Closes #1063 from wesm/ARROW-1156 and squashes the following commits:

166d1a50 [Wes McKinney] iwyu
34f5c9d1 [Wes McKinney] Harden default cast options, fix unsafe Python case
1d07b756 [Wes McKinney] Add some basic casting unit tests in Python
c1b45709 [Wes McKinney] Expose arrow::compute::Cast in Python as Array.cast. Still need to write tests
a9a04c9c [Wes McKinney] UnaryKernel::Call returns Status for now for simplicity. Support pre-allocated memory
8903709b [Wes McKinney] Implement casts from null to numbers. Try to cast for types where we do not have an inference rule when converting from arrays of Python objects
a22dd20a [Wes McKinney] Add test to assert zero copy for compatible integer to date/time
a14b83f7 [Wes McKinney] Create callable CastKernel object. Add zero-copy cast rules for date/time types