Avoid capturing MArrays in BoundsError

This quirk allows bounds checking of `MArray`s on device in more cases.
This is a workaround to address
<JuliaGPU/CUDA.jl#2313>.

This change was proposed upstream here
   <JuliaGPU/CUDA.jl#2314>
and is no longer needed if accepted upstream.

ext/RavenCUDAExt.jl

@@ -3,6 +3,7 @@ module RavenCUDAExt
 import Raven
 import Adapt
 import MPI
+import StaticArrays
 
 isdefined(Base, :get_extension) ? (using CUDA) : (using ..CUDA)
 isdefined(Base, :get_extension) ? (using CUDA.CUDAKernels) : (using ..CUDA.CUDAKernels)
@@ -58,4 +59,11 @@ else
     Adapt.adapt_storage(::CUDA.KernelAdaptor, ::MPI.Comm) = nothing
 end
 
+CUDA.@device_override function Base.checkbounds(A::StaticArrays.MArray, I...)
+    @inline
+    checkbounds(Bool, A, I...) ||
+        CUDA.@print_and_throw("BoundsError while indexing an MArray.")
+    nothing
+end
+
 end # module RavenCUDAExt