fix(modeling_utils): rename shadowed tuple loop var in get_parameter_dtype (#13789)

[Anai-Guo]

What

Fixes #13789.

Inside get_parameter_dtype (src/diffusers/models/modeling_utils.py), the
inner find_tensor_attributes fallback used for nn.DataParallel
compatibility is annotated with list[tuple[str, Tensor]]. The enclosing
function also did:

for tuple in gen:
    last_tuple = tuple
    if tuple[1].is_floating_point():
        return tuple[1].dtype

Assigning to tuple anywhere in get_parameter_dtype makes it a function-
local for the entire scope. When execution reaches the DataParallel fallback
before the for loop has ever bound tuple — which is exactly what happens
under nn.DataParallel where the replicated module on the non-master device
has empty named_parameters() / named_buffers() — the inner function
annotation evaluates tuple[str, Tensor] and raises:

UnboundLocalError: cannot access local variable tuple where it is not associated with a value

This breaks UNet2DModel(...).cuda(); torch.nn.DataParallel(model, ...)
forward, because unet_2d.UNet2DModel.forward does t_emb.to(dtype=self.dtype).

Fix

Rename the loop variable to named_tensor so the builtin tuple is no
longer shadowed. get_parameter_device already uses a distinct name
(first_tuple) and is not affected.

Minimal diff in get_parameter_dtype:

     gen = parameter._named_members(get_members_fn=find_tensor_attributes)
     last_tuple = None
-    for tuple in gen:
-        last_tuple = tuple
-        if tuple[1].is_floating_point():
-            return tuple[1].dtype
+    for named_tensor in gen:
+        last_tuple = named_tensor
+        if named_tensor[1].is_floating_point():
+            return named_tensor[1].dtype

Repro

The repro from #13789 (two-GPU nn.DataParallel(UNet2DModel(...))) raises
UnboundLocalError on main and runs to completion with this patch.

---

AI-assisted, human reviewed.