[DOC] Dynamic-layout guide example uses incompatible singleton strides #3760

Report incorrect documentation

Location: Framework integration guide, dynamic-layout code example, at the v4.8.1 source revision.

Problem and reproduction: The example assumes that singleton dimensions of b have stride 1, but the shown PyTorch construction retains stride 4:

import torch
from cutlass.cute.runtime import from_dlpack

b = torch.empty(32, 1, 1, 1, 4).permute(3, 4, 1, 0, 2)
print(b.stride())  # (4, 1, 4, 4, 4)
from_dlpack(b).mark_layout_dynamic(leading_dim=0)
# Expected strides[leading_dim] == 1, but got 4

The same mismatch affects the later leading_dim=2 example. The t1 and t2 examples also print t2 and t3, respectively, before those variables have been defined.

Verification: Ran the six positive examples t0–t5 in a GH200 environment with the main editable core and torch 2.11.0, and an H100 environment with CuTe DSL 4.8.1 and torch 2.13.0. The original and a print-only repair fail at the stride check. Using explicit singleton strides and correcting the print variables makes all six layouts match the documented output in both environments. This checks tensor-layout metadata, not GPU kernel performance; deliberate-error snippets elsewhere in the guide are outside this report.

Suggested fix: Construct b with torch.empty_strided((1, 4, 1, 32, 1), (1, 1, 1, 4, 1)), and print the variables just defined.