Add support for make install (still somewhat experimental, but ready to be tested by downstream packagers)

[fingolfin]

I wrote this a year ago and never finished. I figured I should post it here, to (a) remind me about it and (b) see if anybody has feedback, or even (c) is willing to help.

Help could range from trying it out and reporting issues as they are encountered (I am sure there are some, perhaps even many, as I didn't finish this), over suggesting things to actually implementing improvements.

Resolves #197

Some TODOs I do recall are these (and the patch also contains some):

• add a CI test for this feature
  • i.e., make install GAP somewhere, then try to run it, and also try to build some select GAP packages against it
  • also test make install DESTDIR=/tmp/or/whatever to check that DESTDIR support works (see the automake manual for more on DESTDIR)
  • writing a CI test probably should be the first thing to do when continuing this PR, as it'd surely reveal issues, which then could be fixed in tandem with finishing the CI test
• deal with config.h
  • see the FAQ entry in README.buildsys.md for some more details, and feel free to ask me for more if you are interested
    • actually I made a lot of progress on removing need for config.h from most headers; but there are some hold outs: actual configuration choices (see next point); the kernel version (see point after the next); and SYS_IS_64_BIT
    • so we still may want to generate a gap-config.h for things like HPCGAP or USE_GASMAN or USE_JULIA_GC. Then again, ideally we may want to share the same headers between regular GAP, HPC-GAP, Julia GAP, ... which would suggest that it'd be better to add these to GAP_CPPFLAGS (so it'd end up in sysinfo.gap, which would exist separately for each build variant of GAP; and would also be automatically used by gac this way
    • but GAP_KERNEL_MAJOR_VERSION and GAP_KERNEL_MINOR_VERSION should still be common, so perhaps that could be put into a specific header after all...
    • the definition of SYS_IS_64_BIT still relies on SIZEOF_VOID_P. But I worked hard on not using it in headers, and only one use (in intobj.h is left). But of course packages use (cvec, datastructures, NormalizInterface). I guess this is another thing we could (should?) define in GAP_CPPFLAGS.... Or perhaps we can use a trick as in https://stackoverflow.com/a/32717129/928031
  • but for starters, we could just install it "as is", that should be OK for many use cases and surely beats not having make install at all
• deal with gac
  • see PR Rewrite gac to not use libtool #4746
• deal with the case were we are using bundled GMP/Boehm/libatomic/... -- presumably just error then? Or maybe install them? Or ...?
• document make install (including how it interacts with the configure --prefix and DESTDIR -- can be short and point to the automake manual, as this is mostly for packagers who generally are familiar with this)
• ... more... I'll add it as I think of it

[fingolfin]

I've made major progress towards getting rid of the need to installing config.h in PRs #4678 and #4674. With these, this PR now can finally get rid of all includes of config.h in headers!

This mainly leaves gac and of course CI tests.