`Nx.Random.normal` fails when using `{Torchx.Backend, device: :mps}` as reported in #1608

[VasiliyS]

Attempt ot use Nx.Random.normal or Nx.Random.uniform with the {Torchx.Backend, device: :mps} set as default fails.

There a locked but open issue #1608 that reported the same issue.

Error:

** (ArgumentError) cannot perform operation across devices cpu and mps
    (torchx 0.10.2) lib/torchx.ex:487: anonymous fn/2 in Torchx.prepare_tensors!/1
    (elixir 1.19.3) lib/enum.ex:1814: Enum."-map_reduce/3-lists^mapfoldl/2-0-"/3
    (elixir 1.19.3) lib/enum.ex:1814: Enum."-map_reduce/3-lists^mapfoldl/2-0-"/3
    (torchx 0.10.2) lib/torchx.ex:286: Torchx.left_shift/2
    (torchx 0.10.2) lib/torchx/backend.ex:709: Torchx.Backend.left_shift/3
    (nx 0.10.0) lib/nx/defn/evaluator.ex:461: Nx.Defn.Evaluator.eval_apply/4
    (nx 0.10.0) lib/nx/defn/evaluator.ex:274: Nx.Defn.Evaluator.eval/3
    #cell:kqxskh7zrfmayiam:2: (file)

Looks like the exception starts at the call to Nx.Random.split.

[josevalim]

I closed that one to have this as source. In any case, the root cause is #679.

[pxp9]

Should #1652 PR close this issue ?