Use unified memory for array allocations.

[maleadt]

Massively speeds up host operations, without losing assertscalar-like functionality.

Doesn't properly works right now, due to the coherency requirements of unified memory on pre-sm_60 hardware. Basically, we'd need to synchronize the device before doing any load from unified memory to make sure there's no kernel running (which claims ownership of the unified memory). That would wreck performance again.

I had hoped that JuliaGPU/CUDAdrv.jl#141 would fix this, but that only works around the issue when doing explicit copies (doing a cudaMemcpy instead of just loading from the memory). When indexing, we of course still do a regular memory load, running into the above issue.

If and when we solve this problem, host/unified<->unified transfers should probably do a regular memcpy again (ie. what JuliaGPU/CUDAdrv.jl#141 reverted, but for good reasons).