With this it is now possible to do

foobar = executable('foobar', ...)
meson.override_find_program('foobar', foobar)

Which is convenient for a project like protobuf which produces both a
dependency and a tool. If protobuf is updated to use
override_find_program, it can be used as

protobuf_dep = dependency('protobuf', version : '>=3.3.1',
                          fallback : ['protobuf', 'protobuf_dep'])
protoc_prog = find_program('protoc')

* Don't use __str__.
* Use ExternalProgramHolder for holding an executable.
* Add tests of the dependencies.
* Mark executable in test as native.
* Error if an executable is used with run_command.

I went with an error on the last item because I think that there is no
non error behavior that is always correct.

If a program is not available on the system, we would need an error
one way or the other.

The interesting case then is when the program is available on the
system but is overridden anyway. I would assume that this would
normally happens because the one available on the system is too old.

Should we use one version during configuration and another one during
the build? That sounds dangerous as there is no guarantee that the two
versions are compatible.

It also seems inconsistent with the fact that we error if a program is
overridden after a find_program (see test cases/failing/73 override
used).

BTW, is there any way to test which error caused a test in the failing
subdirectory to fail?

* Improve function name and add docstring.
* Improve error message to say which program was overridden by which
  executable.