Unfortunately necessary because Ion Fury savegames store 120 GB(!!) of data, mostly zeros.
Unlike the old method, this compresses the entire savegame as one block using a ZLib stream so it should be a lot more efficient now.
This broke the savegame reader which still assumed it was working on compressed data. Everything will now take the uncompressed path.
In-stream optional compression is not a good idea anyway, this can and should be done better.
Also: Why is the savegame format architecture dependent???
This is still just glue code, the Build implementation suffers from poor support for multiple keys bound to a single button so they are due for replacement as soon as input is working again.
* moved the ASCII conversion hackery in SDLayer to a subfunction because this made things just messy.
* integrated the keyboard callback's functionality directly into inputState for consolidation purposes. This was yet another independent layer in the keyboard management.
* hook up D_PostEvent as the central place to dispatch keyboard input. This is now the only function that is getting called from the backend and a major prerequiside for swapping out the backend for GZDoom's.
Todo: Route mouse input through that, too.
Unfortunately this means that the keybinding menus in all games except Blood are shot to shit right now because of how they passed the data on to its destination.
These menus are not fixable, this will have to wait until the replacement is up.
For main resource data this is probably unnecessary - most resources are never cached with the exception of sounds and textures, which are loaded permanently anyway.
But for hardware textures this is different. Due to the poor precaching it is impossible to selectively evict hardware textures that are not needed any longer, so for this an MRU cache is really needed so that they do not accumulate and congest the video RAM in the process.
That's one more third party dependency down.
Not only are two hashing algorithms redundant, there was also a large size discrepancy: SuperFastHash is 3 kb of source code while xxhash is 120kb and generally extremely awful code.
It was easy to make a choice here. None of the use cases require this kind of performance tweaking, the longest hashed block of data is a 768 byte palette.
Unfortunately this means that looking up ID 0 can be a bit more costly than the rest because all ID-less entries in RFF files also use 0.
For other file types -1 is used.
