Cache the transformers model used in ONNX test (pytorch#96793)

Also updating merge_rule to allow ONNX exporter team to update the Docker script by themselves.  By default, the model is cached at ~/.cache/huggingface/hub/ under CI jenkins user.

The model is cached so that we don't need to re-download it every time in CI, which causes flaky [CI failures](https://hud.pytorch.org/failure/FAILED%20test%2Fonnx%2Ftest_fx_to_onnx_with_onnxruntime.py%3A%3ATestFxToOnnxWithOnnxRuntime%3A%3Atest_large_scale_exporter_with_tiny_gpt2%20-%20requests.exceptions.ReadTimeout%3A%20HTTPSConnectionPool(host%3D'huggingface.co'%2C%20port%3D443)%3A%20Read%20timed%20out.%20(read%20timeout%3D10.0)).

This is the second part after pytorch#96590

### Testing

Confirm that the model is cached in the Docker image before running the test:

```
jenkins@dd0db85dd34f:~/workspace$ ls -la ~/.cache/huggingface/hub/models--sshleifer--tiny-gpt2/*
/var/lib/jenkins/.cache/huggingface/hub/models--sshleifer--tiny-gpt2/blobs:
total 2460
drwxrwxr-x 2 jenkins jenkins     126 Mar 15 05:48 .
drwxrwxr-x 5 jenkins jenkins      48 Mar 15 05:48 ..
-rw-rw-r-- 1 jenkins jenkins     662 Mar 15 05:48 2c81a6c4c984e95a45338c64a7445c1f0f88077f
-rw-rw-r-- 1 jenkins jenkins 2514146 Mar 15 05:48 b706b24034032bdfe765ded5ab6403d201d295a995b790cb24c74becca5c04e6

/var/lib/jenkins/.cache/huggingface/hub/models--sshleifer--tiny-gpt2/refs:
total 4
drwxrwxr-x 2 jenkins jenkins 18 Mar 15 05:48 .
drwxrwxr-x 5 jenkins jenkins 48 Mar 15 05:48 ..
-rw-rw-r-- 1 jenkins jenkins 40 Mar 15 05:48 main

/var/lib/jenkins/.cache/huggingface/hub/models--sshleifer--tiny-gpt2/snapshots:
total 0
drwxrwxr-x 3 jenkins jenkins 54 Mar 15 05:48 .
drwxrwxr-x 5 jenkins jenkins 48 Mar 15 05:48 ..
drwxrwxr-x 2 jenkins jenkins 50 Mar 15 05:48 5f91d94bd9cd7190a9f3216ff93cd1dd95f2c7be
```

Pull Request resolved: pytorch#96793
Approved by: https://github.com/titaiwangms, https://github.com/ZainRizvi

Author: huydhn

.ci/docker/common/install_onnx.sh

@@ -23,3 +23,18 @@ pip_install \
 
 # TODO: change this when onnx-script is on testPypi
 pip_install "onnx-script@git+https://github.com/microsoft/onnx-script@29241e15f5182be1384f1cf6ba203d7e2e125196"
+
+# Cache the transformers model to be used later by ONNX tests. We need to run the transformers
+# package to download the model. By default, the model is cached at ~/.cache/huggingface/hub/
+IMPORT_SCRIPT_FILENAME="/tmp/onnx_import_script.py"
+as_jenkins echo 'import transformers; transformers.AutoModel.from_pretrained("sshleifer/tiny-gpt2");' > "${IMPORT_SCRIPT_FILENAME}"
+
+# Need a PyTorch version for transformers to work
+pip_install --pre torch --index-url https://download.pytorch.org/whl/nightly/cpu
+# Very weird quoting behavior here https://github.com/conda/conda/issues/10972,
+# so echo the command to a file and run the file instead
+conda_run python "${IMPORT_SCRIPT_FILENAME}"
+
+# Cleaning up
+conda_run pip uninstall -y torch
+rm "${IMPORT_SCRIPT_FILENAME}" || true

.github/merge_rules.yaml

@@ -2,6 +2,7 @@
   patterns:
   - .ci/caffe2/*
   - .ci/onnx/*
+  - .ci/docker/common/install_onnx.sh
   - aten/src/ATen/core/interned_strings.h
   - docs/source/onnx.rst
   - docs/source/onnx*