[dynamo bug burndown] update tensor creation to support sequences of tensors (#120872)

Fixes https://github.com/pytorch/pytorch/issues/120645

`_internal_new_from_data` calls `_recursive_build`, but we run into an error such as the cases.
```
Failed running call_function <function tensor at 0xDEADBEEF>:
scalar_tensor(): argument (position 1) must be Number, not FakeTensor

# e.g. cases
1. [FakeTensor(..., size=(20, 1), dtype=torch.float64), ..., FakeTensor(..., size=(20, 1), dtype=torch.float64)]
- Here, we call _recursive_build(sizes=[4] ...) which hits the base case `if dim == ndim:` in the 2nd level of recursion.
- So, we try to return `scalar_tensor(FakeTensor)`
2. [[(FakeTensor(..., size=(1,), dtype=torch.int64), FakeTensor(..., size=(), dtype=torch.int64)]]

# site note: when can size = ()? Probably from scalar_tensor.
>>> torch.scalar_tensor(1).shape
torch.Size([])
```

Pull Request resolved: https://github.com/pytorch/pytorch/pull/120872
Approved by: https://github.com/ezyang

torch/_refs/__init__.py

@@ -6332,21 +6332,15 @@ def _infer_scalar_type(obj):
 
 # Analogous to recursive_store
 # xref: recursive_store in torch/csrc/utils/tensor_new.cpp
-def _recursive_build(sizes, dim, scalarType, obj):
+def _recursive_build(scalarType: torch.dtype, obj: TensorOrNumberLikeType):
-    ndim = len(sizes)
+    if isinstance(obj, Tensor) and obj.ndim <= 1:
-    assert dim <= ndim
+        obj = obj.item()
-    if dim == ndim:
+        # fall through into next case
+    if isinstance(obj, Number):
         return torch.scalar_tensor(obj, dtype=scalarType)
-    n = sizes[dim]
+
     seq = obj
-    seq_size = len(seq)
+    return torch.stack([_recursive_build(scalarType, item) for item in seq])
-    if seq_size != n:
-        raise ValueError(
-            f"expected sequence of length {n} at dim {dim} (got {seq_size})"
-        )
-    return torch.stack(
-        [_recursive_build(sizes, dim + 1, scalarType, item) for item in seq]
-    )
 
 
 # xref: internal_new_from_data in torch/csrc/utils/tensor_new.cpp
@@ -6383,7 +6377,6 @@ def _internal_new_from_data(
     # TODO: test for numpy input with PyArray_Check
 
     device = device_opt if device_opt is not None else options["device"]
-    sizes = _compute_sizes(data, scalar_type)
     inferred_scalar_type = _infer_scalar_type(data) if type_inference else scalar_type
 
     # NB: Don't need to avoid tracing, as we aren't going to do any manual
@@ -6398,7 +6391,7 @@ def _internal_new_from_data(
         # of a freshly allocated CPU tensor.  Here, we're going to do an
         # alternate, heinously slow implementation: turn each individual
         # scalar into a tensor, and then repeatedly cat them together
-        tensor = _recursive_build(sizes, 0, inferred_scalar_type, data)
+        tensor = _recursive_build(inferred_scalar_type, data)
 
         tensor = tensor.to(device, inferred_scalar_type, non_blocking=False, copy=False)