[QNN] Fix DiT checkpoint loading (#23718)

### Summary

The QNN DiT export test fails while loading
`microsoft/dit-base-finetuned-rvlcdip`: an upstream model-card update
made Transformers reject the existing safetensors conversion, and
requesting a fresh conversion returns HTTP 405 and raises `KeyError:
'event_id'`. See the [failing main
job](https://github.com/pytorch/executorch/actions/runs/38305497935/job/114998525268).

Set `use_safetensors=False` to load the model's published
`pytorch_model.bin` directly, removing the synchronous conversion
dependency from model loading.

### Test plan

With PyTorch `2.13.0+cpu` and CI's pinned Transformers `5.0.0rc1`, the
real checkpoint loads and CPU inference produces finite logits with
shape `(1, 16)`. The loading check can be repeated with:

```bash
DISABLE_SAFETENSORS_CONVERSION=true python - <<'PY'
import torch
from transformers import AutoModelForImageClassification

model = AutoModelForImageClassification.from_pretrained(
    "microsoft/dit-base-finetuned-rvlcdip", use_safetensors=False
).eval()
with torch.inference_mode():
    logits = model(torch.rand(1, 3, 224, 224)).logits
assert logits.shape == (1, 16)
assert torch.isfinite(logits).all()
print(logits.shape)
PY
```

Also verified that loading succeeds with the default
background-conversion settings when `spawn_conversion` is mocked to
raise the same `KeyError`. `python -m black --check
examples/qualcomm/oss_scripts/dit.py` and `git diff --check` passed.
Full QNN export was not rerun locally.

Authored with AI assistance using OpenAI Codex.


cc @cbilgin @psiddh

Author: rascani

examples/qualcomm/oss_scripts/dit.py

@@ -71,7 +71,8 @@ def main(args):
 
     module = (
         AutoModelForImageClassification.from_pretrained(
-            "microsoft/dit-base-finetuned-rvlcdip"
+            "microsoft/dit-base-finetuned-rvlcdip",
+            use_safetensors=False,
         )
         .eval()
         .to("cpu")