Fix DS-CNN Cortex-M input memory format (pytorch#21485)

### Summary

The shared DS-CNN model wrapper returned a contiguous NCHW example
input, while the Cortex-M backend requires channels-last inputs. This
caused the nightly Cortex-M FVP run to abort in
`cortex_m::quantized_depthwise_conv2d.out` with `input must be
channels_last`.

Convert the example input to channels-last, matching the existing
Cortex-M DS-CNN tests and backend contract. The tensor values and eager
model output are unchanged.

### Test plan

Ran `python -m py_compile examples/models/mlperf_tiny/ds_cnn.py` and
`git diff --check`. Loaded the patched wrapper and verified input stride
`(490, 1, 10, 1)`, eager output shape `(1, 12)`, and the exported
user-input stride `(490, 1, 10, 1)`.

Authored with Claude.

Author: JacobSzwejbka

examples/models/mlperf_tiny/ds_cnn.py

@@ -83,4 +83,6 @@ def get_eager_model(self) -> torch.nn.Module:
         return DSCNNKWS().eval()
 
     def get_example_inputs(self):
-        return (torch.rand(1, 1, 49, 10) * 2 - 1,)
+        return (
+            (torch.rand(1, 1, 49, 10) * 2 - 1).to(memory_format=torch.channels_last),
+        )