We get rid of '_is_herm_conj' in favour of '_is_hermitian', replace
'_check_ham' with '_check_hams' (which works on vectorized values)
and add '_check_onsites' (which also works on vectorized values).
Replace absolute value calculation with call to 'cabs' from 'complex.h'

This includes the following changes:

+ Group the elements of 'where' according to the term/site-families
  they belong to. This is pre-computed and stored in '_terms'.
  What specifically is stored depends on whether vectorization
  is enabled.
+ BlockSparseMatrix now takes its matrix elements as a sequence
  of pairs 'which, data', where 'which' indexes 'where'. This
  leverages the values stored in '_terms'.
+ Normalize 'onsite' to a function that takes 'site_range' (int)
  and 'site_offsets' (array of int). This facilitates vectorization
  when it is enabled. Previously 'onsite' was normalized to a
function
  taking 'site' (int), which was not vectorizable.
+ '_eval_onsite' is modified to compose components of '_terms' with
  'onsite'.
+ '_eval_hamiltonian' is modified to compose compoenents of '_terms'
  with 'hamiltonian_term' when vectorized, or 'hamiltonian' when
  not vectorized.

It is not clear how this will be implemented right now.

The next commit(s) will implement the necessary changes to the
operator module to make this test pass.

We will need this normalization in several places, and it is
more general than *just* Hamiltonian terms.

This is the canonical way of checking if a system exposes vectorization
capabilities.

One less import, and the explicit pattern match is not any less readable

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.

We forgot to add him when merging kwant/kwant!333

Require that norbs are specified for abstract vectorized systems.
This was already a de-facto requirement for finalized Builders, and
it should be part of the abstract base class.

Closes #336

See merge request kwant/kwant!334

All site arrays must have 'norbs' set by definition.

The docstring erroneously stated that 'site_ranges' could be 'None',
however we want to restrict VectorizedSystems to contain only site
families that have 'norbs' specified. The docstring is now explicit
about that fact.

fix add_peierls_phase for system without leads

Closes #331

See merge request !333

Incraese the timeout and number of retries when running the
link check on the documentation; this should increase the
specificity of the test.

See merge request kwant/kwant!332
Closes #332.

The combination of longer timeout and larger number of retries should
make the linkcheck sufficiently specific again.

Retrying > once should increase the specificity of the check a lot.

vectorize systems so that value functions may be called once
on a whole vector of sites (for onsite value functions) or a pair of
vectors of sites (for hopping value functions). In this way we avoid
the Python overhead of calling the value functions many times. For systems
with vectorization enabled this can result in a ~100x speedup when evaluating
the Hamiltonian.

See merge request kwant/kwant!323
Closes kwant/kwant#302

Update internal uses of 'hamiltonian' to suppress this warning
and add a TODO to update to the new API.

We now have *System and *System2, so the old 'isinstance' checks
will not be complete. We add convenience functions in 'kwant.system'
and 'kwant.builder' to check against all the relvant classes.
We leave the original 'isinstance' checks in 'kwant.physics.gauge'
because that module does not work with vectorize systems yet.

These will be used in vectorized systems later.

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.

Now 'which' and 'act' (and, by extension, 'to_fd') now accept
SiteArrays as well as Sites.

We remove the ABC metaclass as now 'normalize_tag' and 'normalize_tags'
are cyclically defined, and subclasses must redefine at least 1.