Discretize

Clean up Tree Mesh imports
- pr 63. commits from lheagy, review from rowanc1

Try / except the import of TreeMesh in Tests. This should remove all dependence that the cython code be properly set up in order to import discretize (continues to be an issue on windows...)

IO: UBC meshes
contributions from: thast, micmitch, review from lheagy
from pr 54

- discretize.TensorMesh.readUBC is now a wrapper for loading either 2D or 3D mesh
- discretize.TensorMesh.readUBC can take care of recognizing if the file is a 2D or 3D UBC mesh, or it can be an input given by the user (arg: *meshdim*)
- the original functions are now *hidden* class function under discretize.TensorMesh._readUBC_3DMesh and discretize.TensorMesh._readUBC_2DMesh
- comments in UBC files are assumed to be marked with "!" in front

utils.Zero: Numpy upgrades
contributions from: lheagy, review and comments from rowanc1, jcapriot
from pr 61
- Updates to `Identity()` and `Zero()` so that they are consistent with numpy upgrades in numpy=1.13.0 (https://github.com/charris/numpy/blob/master/doc/source/reference/arrays.classes.rst)
- use scipy.sparse.dia_matrix for spzeros (closed 60)

Noteworthy Changes
- use a `try` / `except` pair when importing the TreeMesh in the main init
- have efficiency warnings state that the user should use `python setup.py install` if cython code is not there in the interp utils
- use a `try` / `except` in the setup.py when importing numpy to tell the user to install numpy if it fails.

Minor Changes
- add codacy, quantified code badges to docs
- pinning to sphinx==1.5.6 for sphinx gallery to build. This is stated in 57 and will be released when upstream changes are made

Minor Updates
- bug fix in MeshIO for UBC mesh (and associated test)
- deploy to pypi from appveyor, travis on tag
- cache the averaging operator

Contributors
- grosenkj
- lheagy
- rowanc1

Interpolation Improvements

implementing the bisection algorithm without the GIL. Gets rid of a lot of python overhead for speed-ups. If using Cython, might as well use it to get down to C-like code.

All of the `for` loops are executed without any calls to the python interpreter, and the bisection algorithm is equivalent to the np.searchsorted without the python overhead. Also, since we know how big the `inds` and `vals` array will be, might as well pre-allocate them.

At least 2x-3x speed on decent sized arrays.

Should address 23

Ints
- bug fix to ensure we are using ints as indices and to create numpy arrays

Utils
- Zero and Identity have a transpose

- include *.pyx files in the MANIFEST.in so they get included in the pypi distribution