Now using debian8, debian9, debian11 for cuda 8 - 11.4

Since cudatoolkit=11.1 isn't in the main anaconda channel (yet?), the nvidia conda channel is used.

We now only build a single package built with numpy 1.11.x for numpy>=1.11,<2 .

There is no guarantee that every future numpy version covered by this
will be compatible with numpy 1.11, but numpy checks this internally at
run-time, and should report this (noisily) when this occurs.

For both libastra and astra-toolbox:

1) We do not use script_env to set CC/CXX anymore, since the compilers
   are installed by conda.
2) The build string is made useful by including either the python+numpy
   version or the cudatoolkit version that the package was built with.
3) Some clean-up of build.sh in buildenv/

For libastra:

1) The libastra.so is built with the conda C/C++ compiler
   toolchain. This has two benefits:
   1) The rpath of libastra.so is set to $ORIGIN, which makes linking
      easier for dependent packages.
   2) libastra.so is linkable against ancient versions of glibc. With
      old versions of memcpy.
2) The C/C++ compiler version is fixed to 5.4.0
3) In libastra/build.sh, we rename $CONDA_PREFIX to
   $PREFIX. Apparently, this is how it is supposed to be done. For me,
   $CONDA_PREFIX was suddenly undefined. Why this was not a problem
   before, is unclear to me.
4) The cudatoolkit runtime dependency is pinned with pin_compatible
5) The libastra conda package now provides headers and .pc file. This
   is useful for building C++ packages that depend on astra.
6) Remove some old code related to cudatoolkit<8.0.

For astra-toolbox:

1) astra-toolbox uses the conda-provided compilers
2) The compilers are fixed to version 7.3
3) Add boost to host requirements of astra-toolbox

Notes on testing:

- The libastra build has been tested with all versions of cudatoolkit
- The astra-toolbox build has been tested with all provided versions
  of python after building a single cudatoolkit version of libastra.

How to test this branch:

- It should work by just editing
  `python/conda/linux_release/buildenv/build.sh`. Set
  BRANCH=CI-use-conda-c-compiler-toolchain
  URL=https://github.com/ahendriksen/astra-toolbox

  and run release.sh from the `python/conda/linux_release` directory.

This fixes the issue where cudatoolkit=8.0 would not install any more
with recent versions of conda. Also, no corruption appears to take
place while downloading packages.

Debian 7 is EOL, and CUDA 10.1 doesn't support its version of glibc.

Hardcoded conda=4.6.14 for now, since 4.7.5 seems to be downloading
corrupted packages when running in docker/linux-64.