Reduce CI model test matrix runtime

[ochougul]

Summary

• add QEFF_TEST_SCOPE=pr|exhaustive and wire Jenkins to default PR runs to pr
• route slow parametrized model suites through a shared test matrix selector
• shrink model/diffusion test configs to tiny valid shapes while preserving architecture/feature coverage keys
• keep exhaustive scope available for full configured model lists

Validation

• pip install -e . using ~/envs/main
• ruff check <changed-python-files>
• ruff format --check <changed-python-files>
• pytest tests/unit_test/models/test_model_quickcheck.py -k 'model_matrix' -q
• QEFF_TEST_SCOPE=pr pytest tests/transformers/models/causal_lm_models tests/transformers/caching tests/transformers/sampler tests/transformers/spd tests/transformers/subfunction -m '((llm_model or feature) and not qnn) and (not full_layers) and (not few_layers)' --collect-only -q
• QEFF_TEST_SCOPE=pr pytest tests/transformers/models/image_text_to_text tests/transformers/models/reranker tests/transformers/models/embedding_models/test_qwen3vl_embedding_mad.py -m '(multimodal) and (not qnn) and (not full_layers) and (not few_layers)' --collect-only -q
• QEFF_TEST_SCOPE=pr pytest tests/diffusers -m 'diffusion_models' --collect-only -q
• QEFF_TEST_SCOPE=pr pytest tests -m '(not on_qaic) and (not finetune) and (not full_layers) and (not few_layers)' --ignore tests/vllm --ignore tests/unit_test --ignore tests/nightly_pipeline --collect-only -q

Notes

• AI assistance used for implementation and validation command execution.
• Human review should confirm the tiny configs still represent the intended architecture/feature paths before merge.

[vbaddi]

per pr fast lane is being is worked here: #1075 and #1094