I used the more modern PEP 518 (pyproject.toml) packaging standard
rather than setuptools, requirements.txt, etc. directly.

Instead of using Cython.Build.cythonize and distuils.Extensions in
setup.py, a typical method to create cython extensions, I opted to use
the scikit-build package, which handles gluing setup.py with CMake.

In particular, some changes to the directory structure, imports and
Cython modules themselves needed to be made to work with scikit-build
because it uses the cython executable rather than the
Cython.Build.cythonize API. This seems a net positive anyway because one
can reproduce the build steps manually for debugging purposes with:

  cython -3 --cplus -I .. -I .. file.pyx

Development can be done in the traditional Python way with some
adjustments for the underlying CMake build system:

  python setup.py build_ext --inplace -- -DCMAKE_OPTIONS=...
  # or
  export GALOIS_CMAKE_ARGS=-DCMAKE_OPTIONS=...
  pip install -e . --no-build-isolation

The python directory structure was reorganized to ease packaging as
well. All Galois Python packages are now under the package galois. The
libgalois_shem bindings are in galois.shmem. To avoid conflicts with
existing CMake target names, the test applications are under
galois._bfs, ... with import aliases galois.bfs, ... respectively.

Need a better way to initialize shared memory runtime.
Using explicit "new" as Default constructors are not being
called.

- Rather than create new symbols like USE_GPU follow the pattern of other
  libgpu/libdist which uses ENABLE_HETRO_GALOIS
  (-D__GALOIS_HET_CUDA__=1) to signal GPU configurations.

- Namespace MAX_SIZE macro to avoid clobbering MAX_SIZE symbols defined
  elsewhere.

- Depend on libraries to provide their include directories and other
  configuration rather than obsolete app function.

- Move third-party dependencies to distinguished external directory
  rather than burying them in our own code.

- Delete unreferenced and non-building code.

Without prefixes it is difficult to determine where to find the source
of includes. Also it makes installation difficult because there there
may be filename conflicts among installations (e.g., util.h vs util.h).

Add "pangolin/" as a directory prefix to all pangolin includes. I didn't
go further to distinguish GPU and CPU versions of pangolin; though that
is an obvious next step. The pangolin library relies on include tricks
to work so defer further refactoring until the library is more separated
from its applications.

Survey propagation is not well-maintained and is racy, which leads to
poor correctness behavior with high concurrency (e.g., increased error,
divergence).

icc fails to compile gtuple due to errors like the following:

  [...]gtuple.h(77): error: pack expansion does not make use of any argument packs
      using append = integer_seq<T, Is..., I>;
                                      ^
            detected during instantiation of class "galois::integer_seq<T, Is...> [with T=int, Is=<>]" at line 104

Similar errors are reported by icc when it compiles <tuple> from the GCC
9 STL library. Currently, icc 12 only supports GCC 8 as a backend.

So, it appears that there are some C++17 features that still trip up
icc. As a workaround, reformulate gtuple in terms of standard C++17 STL
features. Since, gtuple itself was designed to provide transitional
support for integer sequences and related functionality in mixed C++11
codebases, this gives us an opportunity to remove custom code.

icc is very verbose about mismatches between inline, noinline and
implicitly inlined methods (like method definitions in class
definitions). Disable these warnings.

(Apparently, we are not the only ones to do so [1])

[1] https://eigen.tuxfamily.org/dox/DisableStupidWarnings_8h_source.html

Previously, there were two different classes for reductions. One for
types that are efficiently copyable (GSimpleReducible) and one for types
are are expensive to copy (GBigReducible). STL containers manage to deal
with inefficient-to-copy elements without resorting to completely
separate implementations by using move semantics and adding API
adjustments like emplace. This commit does the same.

Replace the two reducible classes with a single Reducible class.
Reducible tries to use move semantics throughout, which will naturally
fall-back to copying if there is no appropriate specialization.

Remove the subclasses of GBigReducible, which existed to handle updating
container types while avoiding creating large intermediate values.
Instead, rely on the getLocal (previously peek_local) method to access
the current per-thread value and require clients to manipulate the value
directly.

This change does lose the clean abstraction property of the previous
approach but having wrapper classes for each composition of container
types seems unwieldy and difficult for users to understand and extend.

when set, reduces amount of per thread storage for each
thread from default of 32MB per thread; useful for machines with
low RAM or for running on CI where memory is limited + multiple
processes are spawned

allocated

This reverts commit c903e49.

also got rid of mem hack define

This reverts commit 1253ea4.

cannot build locally right now so have to test via CI

…unctions

…d only by assert left