Add DLPack support

[ernestds]

So I've left the Julia files in a separate folder and not really incorporated into the main codebase because I don't know where to place them. So for now this is more of a look at what I did and give me a few pointers because I barely know XLA, Reactant and even Git.

Relevant issue: ##1362

Some comments:

• The implementation is heavily based/copied from Jax.
• I am not one hundred percent following the DLPack.jl from_dlpack function convention because I wasn't sure what the best approach would be, i.e, just use their for example DLPack.from_dlpack(::ArrayType,::MemoryLayout, pyobject) or simply create some rarray_from_dlpack outside of DLPack.jl.
• I've only tested with PyCall but I don't see how it wouldn't work with PythonCall.
• For the CPU stuff instead of a data view buffer (client->CreateViewOfDeviceBuffer) I used client->BufferFromHostBuffer with HostBufferSemantics::kMutableZeroCopy because otherwise the buffer would get copied when being donated (I think, all I know about XLA was by just reading its source code while programming this). From my tests the same thing happens in Jax even though it shouldn't (No error message but failing zero-copy? jax-ml/jax#30228)?
• The whole thing is a bit messy and not very well structured as well as missing a thing or two; I will sort it out after I am sure the code in general is ok. In any case, feel free to ask about any general weirdness or confusing stuff.
• Handling of streams for synchronization in DLPack is still something that is in discussion Handling of streams for new API layer dmlc/dlpack#176 https://dmlc.github.io/dlpack/latest/python_spec.html#syntax-for-data-interchange-with-dlpack. I've left a couple functions related to that that can be maybe used in the future (GetBufferComputeStream, WaitUntilBufferReadyOnStream). I believe they work, but I don't know how to test it properly. Some tests I did:

torch = pyimport("torch")
torch_stream = torch.cuda.Stream()
t = torch.tensor(ones(2,2),device = "cuda")
tr = DLPack.from_dlpack(Reactant.ConcretePJRTArray,t)



function cuda_slow_copy_second_to_first(a)
    for i=1:8000000
    a[1] = cos.(a[1])/i
    end
    
    a[1] += a[2]
    return nothing
end
function reactant_slow_copy_first_to_second_(a)
    @trace for i=1:50000
        a[2:2] = cos.(a[2:2])/i
    end
        
    a[2:2] += a[1:1]
    return nothing
end
py"""def pytorch_slow_copy_first_to_second(a):
    for i in range(100000):
        a[1] = $torch.cos(a[1])/(i+1)
        
    a[1] += a[0]
"""

Reactant.set_default_backend("cuda")
vorg = CuArray(reshape(collect(Float64,1:15),3,5))
v_reac_org = Reactant.to_rarray(vorg)
vreac = Reactant.to_rarray(vorg,)
vt = DLPack.share(vreac, torch.from_dlpack)
vcuda = to_julia(vreac)
reactant_slow_copy_first_to_second = @compile sync=false reactant_slow_copy_first_to_second_(vreac)

stream_reac = (@ccall Reactant.MLIR.API.mlir_c.GetBufferComputeStream(first(vreac.data).buffer.buffer::Ptr{Nothing})::Ptr{Nothing})

# because I don't know how to create a CuStream without going through the struct constructor I do this h8 crime
mutable struct temp
    const handle::CUDA.CUstream
    Base.@atomic valid::Bool

    const ctx::Union{Nothing,CUDA.CuContext}
end
t1 = temp(convert(Ptr{CUDA.CUstream_st},convert(CUDA.CUstream, torch_stream.cuda_stream)), (true), current_context())
t2 = temp(convert(Ptr{CUDA.CUstream_st},convert(CUDA.CUstream, stream_reac)), (true), current_context())
cudajl_stream_from_torch = unsafe_load(convert(Ptr{CUDA.CuStream},pointer_from_objref(t1)))
cudajl_stream_from_reactant = unsafe_load(convert(Ptr{CUDA.CuStream},pointer_from_objref(t2)))
other_cudajl_stream = CUDA.create_stream()

for (stream,name) in ((cudajl_stream_from_torch,"torch stream"),(other_cudajl_stream,"another independent stream"))
    vcuda[:] .= 0
    vcuda[2:2] .= 9

    @cuda stream = stream cuda_slow_copy_second_to_first(vcuda);
    py"""
    with $torch.cuda.stream($torch_stream):
        # C = $torch.mm($vt.T, $vt)
        pytorch_slow_copy_first_to_second($vt)
    """
    println(name)

    println(vcuda)
    println()

end

for (stream,name) in ((cudajl_stream_from_reactant,"reactant stream"),(other_cudajl_stream,"another independent stream"))
    vcuda[:] .= 0
    vcuda[2:2] .= 9

    @cuda stream = stream cuda_slow_copy_second_to_first(vcuda);

    reactant_slow_copy_first_to_second(vreac)
    println(name," b4 sync")
    println(vcuda)
    CUDA.synchronize(stream)
    println(name," after sync")
    println(vcuda)
    println()
end

"
reactant stream b4 sync
[9.000000125 0.0 0.0 0.0 0.0; 2.012517860688443e-5 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0]
reactant stream after sync
[9.000000125 0.0 0.0 0.0 0.0; 9.000020124999995 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0]

another independent stream b4 sync
[1.8356336882806083e-6 0.0 0.0 0.0 0.0; 2.0095251488014837e-5 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0]
another independent stream after sync
[2.1951934491260463e-5 0.0 0.0 0.0 0.0; 2.1826934491260464e-5 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0 0.0]

"

I didn't test the reverse order because for whatever the reason Reactant always seems to sync for me.
@lassepe