Rm-tools

-modified the generation of maxPI and peakRM maps in RMsynth3D so that
it occurs in its own function rather than as part of the file writing
function; this allows it to be run inside a notebook.
-modified RMpeakfit_3D to be more stable when FDF cubes have NaNs present.
-Stokes I fit coefficients no longer become NaNs when negative Stokes I
models are renormalized to a new reference frequency.
-added test for unreasonably large peak RM (compared to computed range)
as this can happen when peak is on edge of spectrum. Changes RM to NaN in
these cases.
-Fixed crash when sigmaAdd calculations received NaNs in inputs.
-added experimental tool to compute QU models from clean components,
requested by Luke Pratley. This currently has a normalization error of some
kind, which I've asked Luke to chase down.
-dropping support for Python3.5, since the QUfitting code has problems with it
and it's already end-of-life.
-improved the calculate_RMSF script to not crash when given width smaller than
first sidelobes. Also corrected it to produce the correct width: previously
when width was specified by user it would be 2x wider than specified.
-fixed problem with do_fitIcube.py which would crash when fed 4D Stokes I cube.
-fixed undesired behaviour in rmsynth1d and rmclean1d where spectra plots
wouldn't be saved when plotting to screen was turned on.

-updated calculate_RMSF to use same theoretical width as rmsynth1D
(the formula from Dickey et al. 2019)
-fixed problem where very small Stokes I model values (~1E-60) would
cause numeric overflow in Q/I, U/I, creating infinities that were
not handled gracefully. Now converting infinities to NaNs.
-fixed peakfitcube not being callable from the command line.
-fixed typos in rmtools_calcRMSF.
-fixed pip installation so that QUfitting models are correctly installed.

-QU-fitting updates from Alec Thomson: sigma-clipping of posteriors,
saving evidence to output file, option to not restart from scratch.
-Fixed bug in chunk reassembly script (assemble_chunks.py) where it would
not figure out the cube size correctly when fed the last chunk.
-Updated RMclean1D to use new derotated angle equation for Observed noise
(previously was using old flawed equation fixed in v1.0.2).
-Fixed problem with handling of degenerate FITS axes in RMclean3D
-Created new tool: RMpeakfit_3D.py,which applies the 1D peak fitting to
each pixel in a 3D FDF cube. Outputs 2D maps for the resulting
parameters.
-New Stokes I fitting. Two changes:
1) Logarithmic polynomial fitting function added. This is essentially
a 5-th order polynomial in log-log space, modified into the
standard power-law with curvature formulation. Added flags to
RMsynth1D, do_QUfitting, and do_fitIcube to let user select fitting
function.
2) Dynamic order selection. If user enters a negative polynomial order,
it will now trying fitting successively higher order polynomials
(starting with 0) until the AIC no longer improves.
-default Stokes I fitting is now 2nd order polynomial
-Plotting code has been made more consistent, in terms of no longer
requiring user to press 'Enter', for all routines.
-QUfitting now has .dat output like that of RMsynth1D.
-All supporting tools now have command-line hooks, which all begin with
'rmtools_'.

-Theoretical estimate for FWHM now uses formula from Dickey et al 2019
instead of Brentjens & de Bruyn 2005, as this is more accurate.
-RMSF fitting improved to be more accurate (previously would over-estimate)
-fitReal option for RMSF fitting now works as intended (previously crashed)

-Minor QUfitting fixes by Alec Thomson.
-Fixed crash when writing 3D cubes where all pixels are NaNs.
-changed output names for the QUfitting multinest files.
-Fixed crashes in RMsynth3D when input cubes are all NaNs.
-Fixed crash when RMsynth3D normalizes output cubes by Stokes I model.

-Minor fixes to new plot-saving code.
-Changed theoretical noise estimation equation: Shinsuke Ideguchi pointed
out that it wasn't correctly computing the error in the fractional
polarization. Modified to normalize the channel errors by the model
Stokes I (errors in the model are not accounted for). This will alter
the noise behaviour for sources with steep Stokes I spectra.