v0.9.2  November 2017:  First version to be published on PyPI
v0.9.3  November 2017:  Some optimisations, without any behaviour changes.
        Calculating the likelihood is now significantly faster for large
        interpolated grids and where flux upper bounds aren't used.  This was
        achieved by avoiding using intermediate arrays in the calculations.
        Marginalising PDFs is now faster for grids with 4 or more dimensions
        due to the use of a cache.
v0.9.4  January 2018:  Allowed the use of linear or cubic interpolation when
        interpolating input model grid flux arrays, with the new keyword option
        "interp_order".  Previously only linear interpolation was available.
v0.9.5  Jan-Feb 2018:  Convert print statements to use python logging, and add
        the "verbosity" argument to NB_Model.__call__.  Some minor changes to
        better support unicode in python 2, and include metadata in .pdf files.
v0.9.6  Mar 2018:  Add "likelihood_lines" keyword option.  In the "best model"
        tables, the "Delta_(SDs)" field was renamed to "Resid_Stds" and now
        provides "(obs - model)/sigma" instead of "(model - obs)/sigma".  The
        resolution of output images was increased.
v0.9.7  April 2018:  Added citation info for the NebulaBayes paper to document-
        ation; clarified that the cubic interpolation option is experimental.
v0.9.8  April 2018:  Fixed some bugs and added associated regression tests to
        the test suite for each bug:
        -  Corrected a bug when using the new "likelihood_lines" feature that
        could cause incorrect likelihood calculations (predictions compared to
        the wrong observations).  The NB_Result attributes "obs_flux_arrs" and
        "obs_flux_err_arrs" are now dicts rather than lists.
          -  The calculation of line ratio priors now uses dereddened observed
        fluxes and errors when deredden=True.  Previously results were incorrect
        when deredden=True and the lines in a ratio had a nontrivial wavelength
        separation.  There are new signatures for "calculate_line_ratio_prior"
        and custom prior callbacks (docs/2_Example-advanced.py was updated).
          -  Calculations for the "best model" table, chi^2, and extinction now
        use the gridpoint defined by the peaks of the marginalised 1D PDFs to
        define the "best model".  Previously the n-D PDF peak was inadvertently
        used.  The calculations for parameter estimates are unchanged.  The
        "best_model" dict on NB_nd_pdf objects now has a "grid_location" key
        and the "Index_of_peak" field was added to parameter estimate tables.
v0.9.9  June 2018 - July 2020:  The default grid_error is now 0.1 instead of
        0.35, which will affect parameter estimates.  The interaction of the
        "verbosity" keyword with the logging level was improved, and "ERROR" is
        now an allowed level.  Stop using deprecated pandas method "set_value".

