Honor the dtype argument of aten::mean.dim before reducing

[Raashish Aggarwal (raashish1601)]

aten_mean_dim applied the dtype argument after ReduceMean, so the mean was computed in the input type and only the result was cast. #3009 fixed this for aten::mean (no dim), but aten::mean.dim still has the old order. For an integer input this truncates:

torch.mean(torch.tensor([[1, 2, 2]]), dim=1, dtype=torch.float32)  # tensor([1.6667])
# exported model returns 1.0 (ReduceMean on int64, then Cast)

and for float inputs it keeps the input precision (e.g. [1e8, 1, -1e8] in float32 with dtype=float64 gives 0 instead of 1/3).

The cast now happens before the reduction, like in aten_mean. The rank check runs before the cast, because in full-graph mode the output of Cast has no static shape.

Testing:

• Added an ops.aten.mean.dim_dtype OpInfo (several shapes and dims, plus the precision-sensitive sample from Honor the dtype argument of aten::mean when exporting without dim #3009) and registered it for aten_mean_dim. It fails on main and passes with the change.
• pytest tests/function_libs/torch_lib/ops_test.py -k mean passes locally (9 passed, 68 subtests), and ruff check / format are clean on the changed files.