Cleanup CUDA implementation a bit

[gonzalobg]

• Refactor all kernels into a generic "parallel for" algorithm
  • Supports grid-stride and block-stride loops, configurable with model flag
  • Handles devices of different sizes via occupancy APIs
• Refactor memory allocation APIs
• Prints more GPU details, in particular, the theoretical peak BW in GB/s of the current device, using the NVML library (which is part of the CUDA Toolkit and always available)
• Fixes 2 bugs:
  • Prints the "order" used to run the benchmarks (e.g. classic vs isolated)
  • Fixes a division by zero bug in the solution checking

[gonzalobg]

This was passing. Seems like this and other PRs are spuriously failing due to some cache issue @tom91136 @tomdeakin

[gonzalobg]

Closing for #202