Improved errors when calling device functions on CPU

[dmillard]

Currently, calling a device function in CPU code causes an ambiguous error and then an infinite hang:

julia> thread_position_in_grid_1d()
JIT session error: Symbols not found: [ _julia.air.thread_position_in_grid.i32 ]
JIT session error: Symbols not found: [ _julia.air.thread_position_in_grid.i32 ]

I realized this when trying to call thread_execution_width(), not realizing it was device-only, but it wasn't clear to me if it wasn't properly wrapped or if something else was going on.

Maybe a clearer error could be raised?

[maleadt]

We could take the @device_functions mechanism from CUDA.jl to make a couple of these common functions error: https://github.com/JuliaGPU/CUDA.jl/blob/868d33ea3067d6bb2403a9cf29d0f1214548314b/src/device/intrinsics/indexing.jl#L64