* bind EVP_R_MEMORY_LIMIT_EXCEEDED and update a test

This will allow OpenSSL 1.1.1 on 32-bit (including our Windows 32-bit
builders) to fail as expected. Technically this isn't a malloc error,
but rather failing because the allocation requested is larger than
32-bits, but raising a MemoryError still seems appropriate

* what you want an endif too?

* Raise MemoryError when backend.derive_scrypt can't malloc enough

* Expose ERR_R_MALLOC_FAILURE and use the reason_match pattern to catch it

* Add test_scrypt_malloc_failure in test_scrypt

* let's see if this passes

* add comment to filippo's blog post about scrypt's params

I hate it when we have to do this.

This removes:

- ERR_get_state which really shouldn't be public API.

- A bunch of functions that are really mostly useful within the library
  to add new errors. NB: I say mostly because they are also useful when
  trying to register a new error library, as osrandom does, but osrandom
  is written in C. Python code is more likely to be consuming errors.

- All function codes but EVP_F_EVP_ENCRYPTFINAL_EX because tests still
  reference it. Per PR #3609, function codes are kind of unstable. This
  finishes that up and cleans up the bindings.

- The "line" versions of querying the error queue, just because no one
  seems to be using them and there's a lot.

- Error-printing functions, which make less sense in Python since you'd
  probably wrap in an exception.

Error codes probably could also do with cleaning, but I've left them
alone for now.

* No more FUNCS/MACROS distinction

* change the docs to not talk about MACROS since they're gone

* remove out of date comment

* We always have EC

* We always have ECDH

* We always have ECDSA

* We always have EC codes

* This can go as well

* And this

* unused import