Lazy initialization doesn't trigger with all API calls

[maleadt]

We only lazily initialize when initializing the task-local state, not when e.g. calling cuDeviceGetCount. That can result in bad error reporting:

julia> using CUDA
julia> devices()
Error showing value of type CUDA.DeviceIterator:
ERROR: UndefVarError: `libcuda` not defined
Stacktrace:
  [1] macro expansion
    @ ~/.julia/packages/CUDA/YIj5X/lib/cudadrv/libcuda.jl:3090 [inlined]
  [2] #151
    @ ~/.julia/packages/CUDA/YIj5X/lib/utils/call.jl:27 [inlined]
  [3] check(f::CUDA.var"#151#152"{Base.RefValue{Int32}})
    @ CUDA ~/.julia/packages/CUDA/YIj5X/lib/cudadrv/libcuda.jl:32
  [4] cuDeviceGetCount(count::Base.RefValue{Int32})
    @ CUDA ~/.julia/packages/CUDA/YIj5X/lib/utils/call.jl:26
  [5] ndevices
    @ ~/.julia/packages/CUDA/YIj5X/lib/cudadrv/devices.jl:159 [inlined]
  [6] length
    @ ~/.julia/packages/CUDA/YIj5X/lib/cudadrv/devices.jl:143 [inlined]