WIP: Sparse kronecker v1

Rewrite in Fortran, using the same algorithm.

Unfortunately this is slower than the C++ version. It seems that it is a
compiler issue. Copying A and D into allocatable arrays improves the
situation (looking at the generated assembly code, it seems that gfortran 10 does
not exploit the “contiguous” attribute of pointers, and generates a more
complex code that does index computations with an arbitrary stride; using an
allocatable array removes this problem). However this is still not enough to be
at par with C++ code.