Improve regularisation documentation and add ADMT demo - #427

* Fix the series of transforms to position and orient the bolometer
  cameras.
* Update geometry parameters such that the lines of sight all follow a
  consistant convention and specify the channel ordering in the
  docstring. This required adding the ability to rotate the sensors
  about the viewing axis to reverse the channel order of the
  horizontal-poloidal camera.
* Correct the docstring of load_bolometers to specify the correct
  number of 4-channel sensors in each camera.
* Expand docstrings and comments to better describe the intricacies of
  setting up the bolometer geometry.

This is done automatically if the derivative operators are
sparse. A dense operator is returned otherwise.

An alternative to `invert_regularised_nnls` using sparse weight and
penalty matrices. The numerical algorithm used is slightly different
to `scipy.optimize.nnls`, as it uses the TRF variant of a bounded
`scipy.optimize.lsq_linear` with the lower bound set to 0 to enforce
positivity. Since the results differ due to floating point precision,
the sparse variant is implemented as a separate function to maintain
backwards compatibility.

* Produce additional inversions which incorporate Generomak's
  tangentially-viewing channels, for comparison with the poloidal-only
  inversions.
* Correctly scale the emissivity.
* Fix removing the ray transfer grid from the world before
  forward-modelling the bolometer measurements.
* Better explain why we're building the 1D <-> 2D maps by hand instead
  of just using the ray transfer matrix capabilities directly: this is
  for generality.
* Use the sparse variant of the NNLS inversion for speed and memory
  efficiency. The grid is large enough for this to be worthwhile.

This makes it easier for users to explore the geometry and potentially
even edit it for testing purposes.

This is a more common term in the literature for a bolometer unit
integrating several (typically 4) channels.

* The 1D-to-2D and 2D-to-1D maps are the inverse of one another, so
  one can be computed from the other. Depending on how the end user
  forms their inversion grid it may be simpler to calculate the
  1D-to-2D or the 2D-to-1D, so allow the caller to pass either and
  compute any missing mapping. If the caller already has both
  mappings (machine packages like cherab-jet, cherab-mastu and
  cherab-aug already generate both) then accept both too.
* Add tests for auto-computing missing mappings.
* Tests uncovered a bug converting the Dyy sparse operator to dense
  when `sparse=False` was passed: fix that and complete test coverage
  for `admt_utils`.
* Simplify the ADMT script by only computing the 1D-to-2D map as this
  is fairly trivially obtained from the Ray Transfer object.

Consisdently use (R, Z) coordinates in the poloidal plane