generate_percentiles test fails on `improver_conda_forge` env

[benowen-bom]

Running the acceptance tests using the conda-forge.yml environment is failing on acceptance/test_generate_percentiles.py, specifically the test_masked_percentiles test; the test output differs from the kgo at a selection of points.

I've tracked down the issue to the fact that fast_interp_same_y and slow_interp_same_y return slightly different results, with the fast variant the preferred option (which was used to generate the data), but the slow variant being used in the conda-forge environment as numba is not available.

A minimal example that highlights the issue is the following:

x = [[0.25, 0.5, 0.75],]
xp = [[0, 0.5, 0.5, 0.833333, 1],]
fp = [[-4000. 0 .152. 15000. 20000.],]

The slow variant returns: [-2000. 152. 11288.001] while the fast variant return [-2000. 0. 11288.001]. The issue comes down to how the values are interpolated on the interval with degenerate probability values.

Acceptance criteria:

• fast_interp_same_y and slow_interp_same_y produce equivalent outputs

[btrotta-bom]

In general, interpolation is not well-defined when xp contains duplicate values, e.g. see the warning here: https://numpy.org/devdocs/reference/generated/numpy.interp.html

A workaround could be to change the code (for both fast and slow methods) to remove duplicate xp values (and corresponding fp values) before interpolating.