Summary

The current interface is neither very intuitive nor practical nor in line with the traditional gridspec logic.

Proposed fix

Half a note to self, because this is not completely thought through:

• The parameter ngrids seems not to be working (usage raises)
• The attributes ngrids is not helpful. It's misnomed (naxes would be better), and it's equivalent to len(grid) or len(grid.axes_all).
• Not sure what we need grid.axes_llc for. That seems quite specific.
• The 2D grid arrangement axes_row, axes_column are lists of lists. In contrast subplots returns an array. Thus, you need double indexing grid.axes_row[row][col} and cannot use the multi-index axs[row, col].
• Having the direction parameter and grid.axes_row, grid.axes_column is basically two redundant approaches towards ordering.

Possible improvements:

• deprecate ngrids attribute - replaced by Fix ngrids support in axes_grid.Grid(). #27972
• add an attribute grid.axes or maybe even grid.axs (in analogy to fig, axs = plt.subplots(2, 3)), which is a ndarray of the Axes. Hence one could do grid.axs[row, column].
• maybe (??) soft-deprecate axes_all, axes_row, axes_col, direction?
• in Grid.__init__ make everything after nrows_ncols kw-only
• would be nice to have a default rect, but that requires reorganizing parameter order because that's before nrows_ncols.
  Might be possible migrate

  def __init__(self, fig, rect, nrows_ncols, *, ...)
  

  to

  def __init__(self, nrows_ncols, *, rect, ...)
  

  with reasonable deprecation warnings and catching inappropriate use. - But would have to think this through