Added performance analysis as a feature with AutoModelForCausalLM

[ochougul]

Summary

• Added evaluate_performance(...) to QEFFAutoModelForCausalLM for end-to-end performance analysis: compile + qaic-runner + qaic-opstats.
• Compile perf flags are always enabled: aic_perf_metrics=True, aic_perf_warning=True; for raw_device_stats, also force stats_level=70, ddr_stats=True, aic_pmu_recipe="KernelUtil".
• Added prefill_only to evaluate_performance(...) and now forward it to compile(...).

Key Behavior Changes

Stage selection is now:

• prefill_only=True -> prefill-only
• prefill_seq_len==1 -> decode-only
• otherwise -> both prefill + decode

Artifacts and Paths

• Standardized output layout: compile/, io/, performance_analysis/.
• Added per-stage subdirs for prefill/decode under io, profiling, runner_outputs, and opstats.

Validation

• Expanded tests in tests/unit_test/utils/test_auto_model_api.py; status: 70 passed.
• Hardware smoke verified both:
• --prefill-only -> only prefill artifacts
• --prompt-len 1 (without --prefill-only) -> only decode artifacts