fix CUDActx overdubs

Author: vchuravy

src/backends/cuda.jl

@@ -201,7 +201,7 @@ end
 ###
 # GPU implementation of shared memory
 ###
-@inline function Cassette.overdub(ctx::CPUCtx, ::typeof(SharedMemory), ::Type{T}, ::Val{Dims}, ::Val{Id}) where {T, Dims, Id}
+@inline function Cassette.overdub(ctx::CUDACtx, ::typeof(SharedMemory), ::Type{T}, ::Val{Dims}, ::Val{Id}) where {T, Dims, Id}
     ptr = CUDAnative._shmem(Val(Id), T, Val(prod(Dims)))
     CUDAnative.CuDeviceArray(Dims, CUDAnative.DevicePtr{T, CUDAnative.AS.Shared}(ptr))
 end
@@ -211,6 +211,6 @@ end
 # - private memory for each workitem
 ###
 
-@inline function Cassette.overdub(ctx::CPUCtx, ::typeof(Scratchpad), ::Type{T}, ::Val{Dims}) where {T, Dims}
+@inline function Cassette.overdub(ctx::CUDACtx, ::typeof(Scratchpad), ::Type{T}, ::Val{Dims}) where {T, Dims}
     MArray{__size(Dims), T}(undef)
 end