QEff library utility to capture environment/runtime/library metadata right after loading a model before any transforms are applied.

[quic-dhirajku]

Added a QEff library utility, wired it into core model and diffusers flows, and validated that dumps are created once per model even when later stages fail.

This captures environment/runtime/library metadata in a robust, non-intrusive way for CausalLM, VLM, and diffusers CPU paths.

• Add QEfficient/utils/library_dump.py with dump_qeff_library_once(model_architecture, model_name) that writes a JSON manifest under QEFF_HOME/<arch_or_name>/<model_name>/ qeff_library-<run_id>.json
• Capture schema version, creation timestamp, model name/architecture, QEff package info (version/import path/editable flag), runtime (python/platform/pid), QAIC SDK versions (apps/platform from qaic-version-util), and a full pip-style dependency snapshot
• Ensure the manifest is written only once per model per process via a process-level _written registry
• Add dump_qeff_library_once into QEFFBaseModel.init before any PyTorch transforms so dumps survive transform/export failures and never break model construction
• Add unit tests in tests/unit_test/utils/test_library_dump.py to cover manifest shape, paths, deduplication, QAIC SDK parsing, and basic resilience
• Add tests/unit_test/utils/test_library_dump_api_calls.py to validate CausalLM, VLM (qwen3_vl custom config), Diffusers (Flux pipeline) with this utility.