Fix eager attention selection during model loading

[fangyangci]

Summary

• select the configured Transformers attention implementation before from_pretrained constructs attention modules
• thread export compatibility through both build and direct export paths
• cover loader and build propagation while keeping export test doubles compatible

Motivation

DistilBERT exports could instantiate DistilBertSdpaAttention and only switch the config to eager immediately before ONNX export. This left the module implementation and config inconsistent, causing integer attention masks to reach SDPA and fail with:

Expected attn_mask dtype to be bool or float or to match query dtype

Validation

• uv run pytest tests/unit/loader/test_load_hf_model.py tests/unit/build/test_hf.py tests/unit/export/test_htp_exporter_attention_compat.py tests/unit/commands/test_config_value_priority.py tests/unit/commands/test_export.py -q --tb=short
• 175 tests passed
• retried all 6 affected QNN GPU models: the attention-mask dtype error was eliminated in all 6; 5 completed perf and accuracy successfully, while one progressed to a separate duplicate metadata_props ONNX validation failure