Refactor the RNG state

[johnlees]

Makes the following changes:

Interface:

• Add the pRNG object back in
• Use a new CUDA error check from @Robadob (https://gist.github.com/Robadob/6a84b47ee450b82efe4a5092b44aa771)
• Use a model per particle (not just one on the Dust object)
• Add an rnorm function
• Add an rnorm buffer (unused)
• Update the rpois function for device use
• Pulls in this patch from dust: https://github.com/mrc-ide/dust/pull/62

Optimisation:

• Try to be more explicit about reading/writing RNG state from global device memory (once at start and end of run)
• Use a contiguous block of memory for RNG state
• Interleave RNG state
• Add some __syncwarp(); statements after branching

Compiling:

• Specify compute versions to compile for (7.0 and 7.5)
• Disable device relocatable code (this was off before, but turned on in some versions)
• Remove std::exchange so C++11 is used instead of C++14 (easier with R)