Plotly does not support arbitrary contour values, but can plot
specific contours if they are an equally-spaced range.  This
commit implements the plotting of contours if the `levels`
keyword argument is passed an `AbstractRange`, or if a set
of arbitrary values are passed.  In the latter case, however,
since this is not supported by Plotly, a range based on the
first and last values of the collection passed in is created
and used to define the contours.  A warning is then issued to
the user.

Otherwise, any other types are assumed to be number-like and
adjusted as previously.  Note that Plotly does not guarantee
the exact number of contours will be used.

This partly addresses JuliaPlots#3356.

See plotly/plotly.js#4503 (Plotly issue
tracking the ability to set arbitrary contours).

- Test that the `levels` keyword argument is correctly stored.
- For the Plotly (and PlotlyJS) backend, test that `levels`
  is correctly converted to backend-specific settings.

The `levels` keyword argument does not actually support
values of type `NTuple{2,Integer}` despite the existing documentation,
so remove the mention of this from `_arg_desc` and update the
description of what `levels` means.

Checking of the `levels` keyword argument will be moved out
of individual backends' code, so we can assume that `levels`
if present in `plotattributes` is either an `AbstractVector` or
`Integer`.

This commit moves the check for the correctness of the `levels`
keyword argument into `RecipesPipeline.preprocess_attributes!(::KW)`
and calls the new function `check_contour_levels`, which in turn
will throw an `ArgumentError` if the argument is not correct.

To that end:
- Remove the check for non-integer and non-vector values of
  `levels` from the PyPlot backend code, and
- Add low-level tests for the new function and high-level tests
  for contour plots.