For the buildsystem now only .d files that exist are included

[fingolfin]

Fixes #4025

Of course ideally we'd also have tests for this... Here is one way we could do that: we could have a Travis test which uses an out-of-tree build where we do make clean before make but after configure. That'll trigger the bug as I fixed it.

We could of course do more thing:

• after the make, remove one file *.d file then run make and verify it is being recreated
• the same but for an .lo file
• add in a "bad" .d file which emulates a leftover from a renamed file (with invalid syntax in it) and verify it is not included (as in: make doesn't run into an error)
• verify that make ; make does not regenerate gap or anything (say, by testing that the second make has no output)
• test mixing an in-tree and an out-of-tree build to verify the OOT build does not accidentally use files from the in-tree build (say by first doing the in-tree build, then replacing all .d and .lo files with garbage, then doing the OOT build -- if it picks up one of the wrong files, it'll run into an error)
• ... more can be added over time. That would have caught some of the regressions I accidentally introduced. Of course it won't catch all problems, but we can at least give it a try

[coveralls]

Coverage remained the same at 84.481% when pulling 2417390 on fingolfin:mh/fix-buildsys into 0b4c569 on gap-system:master.

[fingolfin]

I added a bunch of tests. They involving using the name gen, so there is an indirect conflict with PR #4013 here. I suggest this PR here is merged first (as I think merging PR #4013 will trigger the issue this PR here is meant to fix for everybody working on GAP!). Once this PR here is merged, I'l update PR #4013.

[fingolfin]

Note that the comment already that we "skip any missing ones", but that wasn't true anymore with my previous refactoring of the buildsystem; this PR restores the behavior described in the comment