Add an example for extracting one operation from KleidiCV as well.

As it is known at compilation time whether the streaming or
non-streaming variants should be used.

In some environments it causes issues if the non-streaming routines
are called from streaming-compatible functions. (The correct behaviour
would be that the compiler automatically links to the streaming
variants in such a case.)

Gaussian Blur 15x15 kernels always use the custom sigma variant,
with much better performance and near the same accuracy.
Custom Sigma kernels are uniformized and simplified using
std::reference_wrapper in the SVE variant.

Update documentation with oldest supported GCC version as well.

Benchmarks added for Gaussian blur 21x21 and Medan blur benchmarks are
extended.

NEON variant only. Fixes #9. Root cause:
Due to TryToAvoidTailLoop, when width is bigger than 4 and
is not divisible with 4, it scaled some elements twice.

KleidiCV should not depend on the cpp runtime to let it deploy to C only
environments. The C example's cmake configuration is set up in a way
that CI fails if the developer introduces a runtime dependency on the
cpp standard library.

Make sure that --coverage flag is explicitly added to
CMAKE_EXE_LINKER_FLAGS. The cmake version in CI passed it automaticallly
to the linker command line but it can break on other systems.

There is no need for the multiversioning macro for the API function.

Align the implementation to the current head of OpenCV's 4.x branch.