If a function does not support the requested operation then it must
return quickly to avoid delaying the fallback option.

Changing the order of parameter checks caused some tests to fail, so
the tests have been changed to avoid failing the not-implemented check.

A few corrections to inaccuracies plus some formatting tweaks.

Operations in the Neon backend have both a vector path and a scalar path.
The vector path is used to process most data and the scalar path is used to
process the parts of the data that don't fit into the vector width.
For floating point operations in particular, the results may be very slightly
different between vector and scalar paths.
When using multithreading, images are divided into parts to be processed by
each thread, and this could change which parts of the data end up being
processed by the vector and scalar paths. Since the threading may be
non-deterministic in how it divides up the image, this non-determinism could
leak through in the values of the output. This could cause subtle bugs.

By using C++ lambdas we can avoid a lot of copy-paste.

The behaviour of numeric_limits::min<float> is not what most people expect.
Therefore use lowest instead, as a precautionary measure.

The behaviour of numeric_limits::min<float> is not what most people expect.
Therefore use lowest instead, as a precautionary measure.

std::numeric_limits::min is misleadingly named.
A more accurate name would be min_if_integral_else_min_positive_normalized.
The actual function to give the minimum value is lowest.

The purpose of this change is to avoid the cost of supporting many
versions of OpenCV.

This commit increases the range of values of the input matrix and
the kernels in the Separable Filter 2D conformity tests.

The value "0" is not included to work around a possible OpenCV bug.

This change is required by the latest version of clang-tidy.