At least Automake 1.16.1 does not seem to have problems with this.

This partially reverts "Account for the case when YAXT and PPM are provided as '/path/to/lib/libname.a'."

This enables building with versions of libtool that are patched to avoid
overlinking (e.g. on Debian systems). It is also a bad idea in general to rely
on negative side-effects, which overlinking is, and we try to avoid that.

The conditional indicates that FC is available and all required compiler flags
are known.

The ONLY clause works only with MPI versions that have all MPI subroutines
declared in the module. According to Thomas Jahns, this is not always the case
(see comments in #4 for more details).

This way, we avoid target-specific flags '-DUSE_MPI', which result into
target-specific rules that do not use Libtool for the compilation.

The idea is that when an additional feature is enabled, we should run the same
tests as without it plus additional tests for the feature.

This is the first step towards switching to the parallel test harness.

1. Check for FC only if one of the Fortran interfaces is requested.

2. Fail if the Fortran ISO C interface is enabled but FC is not found.

3. Check for F77 only if the C Fortran interface is requested and either FC was
   not found or the user set F77 explicitly. In the latter case, we make sure
   that the generated library is compatible with both FC and F77.

4. Check for Fortran interfaces of MPI and YAXT only when needed and do not fail
   if either of them not found: we can still provide the C Fortran interface.

5. Refactor the Automake conditionals: remove redundant, rename the required in
   a consistent way. Conditionals from acx_options.m4 left unchanged since they
   might affect the integration into CDO.

6. General clean-up and refactoring.

7. Temporarily drop support for the experimental ruby and python interfaces.

Probably this will need to be revisited.