The removed logic is an artifact from when vectorized and
unvectorized systems shared the same logic for 'hamiltonian_submatrix'.
This was accidentally left in when we reverted to the old logic
for unvectorized systems.

Add the 'subgraphs', 'terms', and 'site_arrays' attributes, and
'hamiltonian_term' method and implement 'hamiltonian_submatrix',
'cell_hamiltonian' and 'inter_cell_hopping' in a vectorized way.

When System.parameters was introduced, we thought it would be a good idea to
also use this new attribute.  The plan was to check in hamiltonian_submatrix()
whether any unneeded parameters are provided by the user.  This turned out to
be impossible, so the only check that remained was whether all the necessary
parameters are indeed provided.  However, this check is not necessary, because
the 'hamiltonian' method is in any way obliged to signal an error when it does
not obtain the necessary parameters.

Additionally update all associated docstrings to say that 'args'
is deprecated in favor of 'params'.

This means we don't need to create a dummy class that implements
__get__ just so we can use hamiltonian_submatrix as a method.

Remove fixes for unsupported versions of numpy, scipy and matplotlib.

Hamiltonian value functions may now have signatures that depend on
different parameters. On finalization, Builders now inspect all of their
value functions and store the names of the parameters on which the value
function depends. Then, when evaluating the Hamiltonian, a particular
value function is only passed the parameters on which it depends.

'System.hamiltonian' has been updated to accept an extra keyword-only
parameter, 'params', which is a dictionary mapping parameter names to
values. This is mutually exclusive with the existing '*args'. All
top-level API that takes an 'args' parameter now also takes a
keyword-only parameter, 'params'

Summary of changes:
* Add a test for trailing whitespace, remove all trailling whitespace
* Add a test for broken URLs in docs, fix all broken URLs and permanent redirects
* Switch sphinx doc build to strict (warnings to errors) and nitpicky (broken links to warnings). Fix all sphinx warnings.
* Deploy docs from any commit to [kwant/kwant](https://gitlab.kwant-project.org/kwant/kwant) to https://test.kwant-project.org/doc/dev
* Deploy docs from any commit to master at kwant/kwant to https://kwant-project.org/doc/dev
* Deploy docs from any tag release tag (so anything that is vX.Y.Z) at kwant/kwant to https://kwant-project.org/doc/X.Y

Under Python 3, `scipy.sparse.coo_matrix` fails when passed memoryviews.
We coerce the memoryviews to numpy arrays before passing them to the
`coo_matrix`. Even though this is kind of a hack it shouldn't be very
inefficient as no data is actually copied. See
https://github.com/scipy/scipy/issues/5123 for details.

Unbound methods are not longer a thing in Python 3 [1].  Instead,
functions use the descriptor protocol [2] to properly bind the `self`
argument if they are being called as a method. Cython functions do not
have a `__get__` attribute and so don't satisfy the descriptor
protocol. They therefore can't be used as methods. As a workaround we
create a wrapper class with the proper __get__ method. The advantage of
this over creating a Python wrapper is that the function signature is
preserved.

[1]: https://docs.python.org/3.0/whatsnew/3.0.html
[2]: https://docs.python.org/3/howto/descriptor.html

This still does not allow to import kwant under other names:
http://thread.gmane.org/gmane.comp.python.cython.user/13809

Once we depend on Python 3.5, and "proper" relative cimports are supported by
Cython, we could add a test for exclusive use of relative imports.

Relative cimports are not supported by Cython so far.  Once this has been
solved we can switch to relative cimports as well and kwant will be importable
under different names.