The compilation of example will still
produce some warnings, but they are
unrelated to the new code

There was a bug in nabla computation
that relied on gray layout

The output type must be floating point,
thus a check was added to make sure it
is the case. Though I had to add
specific cases for float32_t as
std::is_floating_point does not
consider it a floating point type

Now conservation is only within 10%

The test seems to be failing though,
so revisiting the correctness of the
algorithm is needed

Previously algorithm did not
use a buffer image and updated
pixels on the fly, by using
those that were already modified,
which is incorrect. Also reduced
number of warnings in the new files
to zero

Rather than manually
writing out each entry,
use std::transform

The accumulate part was wrong,
it multiplied by delta_t on
every sum, which is wrong

This is just different
discretization of Laplace
operator.
https://en.wikipedia.org/wiki/Discrete_Laplace_operator

This commit implements correct
diffusion and separates the whole
process into multiple pieces

Found a bug inside tests, now it
computes correctly, allowing to reduce
error threshold drastically

Laplace stencils are now the same
as in mathematical notation

The new function will provide
a uniform way to generate stencils
by making sure directions are
indexed properly

The 5 points Laplace stencil is now
adding only required points and not
assuming that others are zero

Since this cannot be captured by ref,
it is captured by value instead.
The change is to not declare
redundant variables

Also decrease seed range to make test
duration manageable.