use correct (sequential) Mumps libraries in testing image

See merge request kwant/kwant!354

Conda forge recently updated their Mumps packaging so that the
sequential version of Mumps is in a separate package,
[mumps-seq](https://anaconda.org/conda-forge/mumps-seq),
as are the include files
[mumps-include](https://anaconda.org/conda-forge/mumps-include).

First steps in ND vectorized systems

Closes #326

See merge request kwant/kwant!341

Following a request by @cwg.

Previously infinite vectorized systems explicitly stored the sites
of the previous unit cell, and one of the tests and 'hamiltonian'
relied upon this fact. We therefore need to change these places
now that infinite vectorized systems store only the sites of the FD.

Because 'sites' and 'id_by_site' are required for backwards
compatibility they must know about the sites of the previous unit cell.
In principle we could compute these as-needed, but explicitly storing
them allows the implementations of _Site and _IdBySite to remain simple.

Previously we checked whether the subgraphs of a term used site
arrays corresponding to sites >= cell_size, now we can directly
check that the symmetry element of a term is not the identity (zero).

+ Store sites in the fundamental domain only
+ Compute and store the symmetry element with each term

Here we continue to build the hopping terms from the graph, however
when extending to handle >1D translational symmetry we will need
to query the Builder directly. We could in principle already do that
at this stage, and it would mean that less has to change later.

Even though we currently import 'Symmetry' and 'NoSymmetry' in
'builder' (to maintain backwards compatibility), we should use
the canonical location within Kwant itself.

N-dimensional systems will need symmetry elements, so the
symmetry definition needs to be in the system module.

Previously we required an array of group elements to act on a
SiteArray (1 element per site), however often we want to act with
a single group element on *all* the sites in a SiteArray.
For TranslationalSymmetry we implement this using array broadcasting.

hamiltonian_submatrix is not meaningful for
infinite systems, and (unlike un-vectorized systems)
is not even used for constructing cell_hamiltonian and
inter_cell_hopping in the vectorized case.

The docstring refers to site orderings, and explicit references to
the graph; these are things that vectorized systems keep for
backwards compatibility, but we wish to move away from. For this
reason we move the docstring to the (unvectorized) infinite system
class and add a more specific docstring to the vectorize infinite
system class.

We now construct the site arrays to pass to 'val' in a single
place (this is more readable).

We had forgotten to pass the 'vectorized' flag when constructing
a Builder with a symmetry

This is a pure refactor.

add plotly plotter

See merge request kwant/kwant!294

Raises a RuntimeError when incompatible inputs are provided to the plot function using the plotly backend.

Called once during setting of plotly backend.