I think python should use the C++ interface rather than the C one.
The reason is that the C++ interface included in PLUMED 2.5
rethrows exceptions and thus can be used to load
a shared library compiled with a different compiler or
linked against a different standard library.

In perspective, this will allow compiling PLUMED with
a different compiler/library than that used to compile
python itself.

now covered by the regtest better

It should prepend the libplumed.so library with prefix -Wl,
so as not to disturb NAMD makefile. (suggested by @jcphill)

Should have been removed with 1221e3e1

cc: @dvdesolve

See #407

As of PLUMED 2.4.1, an error is triggered when aligning using a pdb file
with zeros in the last column. I here fix the provided pdb file.

See this:
https://groups.google.com/d/msgid/plumed-users/4e2f69b8-1b2d-4bd3-a2e3-ecdfeef65d77%40googlegroups.com?utm_medium=email&utm_source=footer

I added a plumed_cmd_nothrow function to the interface, that is used from
C++. This allows better support for mixed compilers (e.g. MD code and PLUMED
compiled with different compilers or linked against different std libraries).

Notice that I tried in the past to achieve the same using cmd("setErrorHandler"),
but this does not work correctly with some combination of compilers.
In addition, it implies throwing from a extern "C" function, which is not standard
and might lead to portability issues.

The plumed_cmd_nothrow implementation catches any exception raised within plumed,
converts it to an error code, and rethrows it within the MD code. This approach
should be more robust.

Set LD_LIBRARY_PATH before running plumed when compiling python
interface

and SCEXP to SCALEINT

todo:
- gpu should only be run by master
- add documentation

disabled arrayfire from travis (but I kept the code if at some point we can make it numerically more stable)

this might decrease measured coverage, but will make the job
faster and possibly capable to finish on travis-ci