Rework main pool alloc to support fragmented buffer. Move zbuffer & fb to different memory banks

[aglab2]

Moves zbuffer and framebuffer to separate memory banks. To implement this, alloc only pool was removed and replaced with main pool instead that supports fragmented memory. Fragmented pool is not using double linked list and is built very similarly to alloc only pool.

Saves 0.5 fps in JRB.

[aglab2]

Updated PR with splitting FBs to each own memory bank with USE_EXT_MEM. Saved another 0.5f

[arthurtilly]

Things that need to be done:

• Make the defines clearer to the average user - ideally slightly less technical and not using magic numbers as each setting. Having fewer settings (only the ones that are necessary) would also be nice.

• Remove config_memory_private.h

• Ensure that puppyprint ram usage displays are still entirely accurate

[aglab2]

Need to add better documentation for main_pool_ functions to clearly state their benefits compared to alloc only pool

[aglab2]

Finished documentation for main pool allocator. I have also profiled level loads - I have implemented fast path for main_pool_alloc that will be inlined to avoid random slowdowns on level startup.

3.0:
castle inside: 103423 105440 106329 102889 103430
castle grounds: 42575 48902 51084 55313 54764
bob: 71654 70644 70339 70760 71486

3.0 new main pool:
castle inside: 106152 106099 104109 104368 103311
castle grounds: 48520 44482 43405 45028 41257
bob: 70940 70981 70684 70690 71391

I also fixed other issues noted in @arthurtilly comment, code is ready for another round of review

[gheskett]

Any chance the z buffer could be moved back to the start of RAM, FB0 to the end of the code segments, and FB1/FB2 bordering 0x80700000? Unless I'm misunderstanding, this would come with the same performance benefits with the least fragmentation necessary. (Map packer loading to 0x80700000 may not be ideal here.)

[aglab2]

I used such combination of memory buffers in one of my hacks, I can add it as an option.