* Put AES ctx on the heap

This forces people to use the ``ctx_release`` functions, because otherwise there will be leaks

* Put fips202 on the heap

* Add much more docs for fips202.h

* fixup! Put fips202 on the heap

* Put SHA2 on the heap-supporting API

* Fix clang-tidy warnings

* Fix unreachable free() in falcon

* Fix McEliece8192128f-sse GNU Makefile

* Add state destroy to SHA2 API

* Include optimized SPHINCS+ implementations

I've generated new implementations from the sphincsplus repository.

* Don't destroy sha256ctx after finalize

* Attempt to shut up MSVC

* Make sure to drop errors in rmtree

* Names passed around (mostly constant) value more clearly
* Put ``const`` where applicable (almost everywhere)
* Renamed ``primitive.h``
* Fixed some nits

This was causing compilation failures with clang, that there were 
duplicate const keywords once the macro was expanded in hash.h

This PR sacrifices passing some extra arguments to get rid of the global
state.

* Haraka needs state in all hash calls, this results in changes to the
  hash functions specified in `hash.h`.  The extra pointers passed would
  not be necessary for SHA256 or SHAKE256.
* SHAKE256 did not have global state, but uniformity in the implementations
  requires us to pass  around the new state context anyway. Otherwise,
  @joostrijneveld's SPHINCS+ generator doesn't really work anymore).

We introduce a new header file called `primitive.h` which defines the
required state type for the generic functions. I did not go into
replacing _all_ occurrences of state variables by the new `hash_state`
macro.

There are schemes, like SABER (#192) that have more than one principal
submitter. Consistency warrants that we turn it into a list for all
schemes and don't do something with allowing either a str or a list:
that would just be very annoying to parse.

Closes #194

Windows already complains about this in CI, but this will
let us catch these issues on Linux as well.

Thanks to Stefan Kölbl