Better integration with Julia GPU stack

[maleadt]

We currently duplicate quite some functionality that could be reused from CUDA.jl (or GPUCompiler.jl) if it were more extensible/reusable:

• Launch syntax (integrate with @cuda or provide something similar)
• Argument conversion during launch (cudaconvert and friends, skipping of ghost values, etc)
• Compilation cache (GPUCompiler.jl's is very much CodeInstance-oriented though, so not sure if this would work, but the current _compilation_cache in cuTile.jl is naive and slow)
• Reflection utilities (hooked compilation etc; maybe not worth the integration effort)