Layered modelling

The simulation class takes new the parameters layered and layered_opts,
where the default values are False and None, respectively. If layered=True,
there will be no 3D computations. Instead, it will create a local layered (1D)
model for each source-receiver pair, and compute the response using the
semi-analytical code empymod (which needs to be installed manually, as it
is a soft dependency). In this case an eventual gradient is computed using the
finite-difference method, not the adjoint-state method, perturbing each layer
slightly. The main purpose of these layered computations is for quick checks,
QC, verifications, etc. Layered computation is also possible through the CLI,
through the new flag -l or --layered, and a new section [layered]
in the config file.

Other changes (many of them related to the above):

• Model instances have a new attribute exctract_1d, which returns a layered
  (1D) model, extracted from the 3D model according the provided parameters;
  see :attr:emg3d.models.Model.extract_1d.

• CLI takes new the boolean add_noise in the section [noise_opts]
  (default is True).

• Maps: New function ellipse_indices returning a boolean indicating which
  points fall within a general ellipse for the provided input parameters.

• Bug fixes, small improvements and maintenance

  • Simulation.misfit returns an ndarray again instead of an DataArray (was
    erroneously changed in v1.2.1).
  • Write json can now handle NumPy int/bool/float.
  • A clean on a Simulation now removes correctly the weights.
  • Capture error in jtvec if weight is complex NaN (should be real).
  • Model: mapping can now be an already instantiated map (sort of
    undocumented).
  • Cleaned-up the namespace by setting dir() explicitly.
  • Replace pytest-flake8 by plain flake8.
  • Moved all multiprocessing-related functions to _multiprocessing.py.