Generic arithmetic codec implementation for CUDA

Data models supported:

• static binary
• adaptive binary
• context-adaptive binary
• context-adaptive byte

Context-adaptive byte model seems to give the best efficiency, while being marginally slower than others. Requires a lot of video memory.