Deterministic code generation

[mstimberg]

Two main issues led to code files changing for identical models (i.e., even when running a model again in a new process where the problems with regard to names as discussed in #680 do not play a role):

1. Variable declarations and other "unordered" code was generated in a different order for every run, and
2. The file names written to the result directory used hash() to deal with the problem of case-insensitivity on Windows (otherwise i and I would be stored to the same file). For security reasons, the hash() function changes its output for every run.

The fix in this PR is to:

1. Use sorted() whenever iterating over variables that are stored in sets, and
2. Use zlib.crc32 instead of hash to calculate the hash value for the filenames. Strictly speaking, crc32 is a checksum and not a hash value and is more prone to lead to collisions, but this should not matter for us since we only need to disambiguate in very specific circumstances (names that are identical except for case).

I did not set up a formal test for this, but I verified with fairly complex examples (e.g. frompapers/Kremer_et_al_2011_barrel_cortex.py) that the generated is identical over runs.

[thesamovar]

Good idea!