Don't assume the model is on a CUDA device

[jpsamaroo]

Currently, DaggerChain communicates to Dagger that the wrapped model is located on a CUDA GPU, which is not necessarily true (and shouldn't be a requirement anyway). We should provide functions which can move the model to the GPU and communicate the correct location to Dagger, and/or auto-detect where a model currently resides.

[DhairyaLGandhi]

DaggerFlux.jl/src/dflux.jl

Line 11 in 101c23b

daglayer(f, args...) = delayed((m,x...) -> m(x...))(Dagger.tochunk(f, DaggerGPU.CuArrayDeviceProc(1, CUDA.device().handle, CUDA.uuid(CUDA.device()))), args...)

is where we assert that something is on the GPU. We can remove the CUDA part out safely now I believe?

[jpsamaroo]

If we remove it, then we won't get automatic GPU execution (because GPU execution is disabled by default). We probably need a dispatch-based API in Dagger to enable GPU execution for certain functions (like DaggerChain).

[DhairyaLGandhi]

Do we then intend to revert the changes to daglayer in #19 before merging?

[jpsamaroo]

I believe this has been resolved.