Fix Spark-X2.5 CI: use writable HF_HOME for model download

The CI runners mount /mnt/hf_cache as read-only. Existing models
(qwen3, phi_4_mini, lfm2_5) are pre-cached there, but Spark-X2.5
models are not, causing hf download to fail with:

  OSError: [Errno 30] Read-only file system: /mnt/hf_cache/hub/models--XHToken--Spark-X2.5-*

Set HF_HOME to a writable temp directory for Spark-X2.5 models only,
following the same pattern used in cuda.yml.

Author: dongjiang1989

.ci/scripts/test_torchao_huggingface_checkpoints.sh

@@ -127,6 +127,8 @@ case "$MODEL_NAME" in
 
   spark_x2_5_1_7b)
     echo "Running Spark-X2.5-1.7B export..."
+    export HF_HOME="${RUNNER_TEMP:-/tmp}/hf_cache"
+    mkdir -p "${HF_HOME}"
     HF_MODEL_DIR=$(hf download XHToken/Spark-X2.5-1.7B)
     EXPECTED_MODEL_SIZE_UPPER_BOUND=$((2 * 1024 * 1024 * 1024)) # 2GB
     $PYTHON_EXECUTABLE -m executorch.examples.models.spark_x2_5.convert_weights \
@@ -151,6 +153,8 @@ case "$MODEL_NAME" in
 
   spark_x2_5_4b)
     echo "Running Spark-X2.5-4B export..."
+    export HF_HOME="${RUNNER_TEMP:-/tmp}/hf_cache"
+    mkdir -p "${HF_HOME}"
     HF_MODEL_DIR=$(hf download XHToken/Spark-X2.5-4B)
     EXPECTED_MODEL_SIZE_UPPER_BOUND=$((5 * 1024 * 1024 * 1024)) # 5GB
     $PYTHON_EXECUTABLE -m executorch.examples.models.spark_x2_5.convert_weights \