Faster (by a magnitude) OpenCL AES code

[magnumripper]

#5591 (comment)

The default KeePass benchmark includes old low iteration test vectors. To see modern, real-life cost, edit opencl_keepass_fmt_plug.c, change the define of KEEPASS_REAL_COST_TEST_VECTORS to 1, rebuild and benchmark with -cost=31250000 (it will take a looong time).

Each KDF-AES round is 32 bytes of AES-256 so at this cost it will encrypt exactly 1 GB with AES-256 per candidate - great for testing AES performance. Our shared OpenCL AES comes in two flavors. Using the "bitsliced AES" (default for GPU - I kinda doubt bitsliced is a correct description but it does some things in parallel), Super's 1080 does 20.5 GB/s which I believe could be improved with at least 10x, (and forcing it to run our older, even worse, vanilla AES it's down to 1.15 GB/s which is essentially useless for anything but decrypting a short verifier).

The faster GPU code (CUDA) is at https://www.github.com/cihangirtezcan/CUDA_AES and there's also a freely available PDF.