Test: Validate Unified Classifier correctly chooses between LoRA path and Traditional path for inference. (#639)

* test: add router path selection E2E tests

Adds comprehensive E2E tests to verify router path selection logic between
LoRA and Traditional inference paths in the Unified Classifier.

Tests verify:
- Small batch (1-3 items) + single task → Traditional path
- Large batch (≥4 items) → LoRA path (18.5% faster)
- Multi-task requests → LoRA path (parallel processing)
- LoRA models loaded via auto-discovery (requires lora_config.json from PR #629)
- Performance comparison shows LoRA efficiency gains

All 5 tests pass, confirming:
- Auto-discovery successfully found LoRA models with lora_config.json
- Router correctly selects paths based on batch size and task count
- LoRA path provides measurable performance improvement (18.5% faster per-item)
- Unified classifier metrics are properly exposed

Related to PR #629 (lora_config.json fix) and demonstrates that the fix
enables proper LoRA model detection and routing.

Signed-off-by: Yossi Ovadia <[email protected]>

* test: clarify 07-router-path-selection focuses on classifier API

Updates documentation in 07-router-path-selection-test.py to clarify:
- Tests direct Classification API (port 8080) only
- Focuses on Unified Classifier's internal path selection logic
- Removed Envoy/ExtProc references from this test
- Added note pointing to future 08-envoy-routing-test.py for full stack tests

No functional changes - all 5 tests still pass:
- test_small_batch_single_task
- test_large_batch_triggers_lora
- test_multi_task_triggers_lora
- test_performance_comparison (LoRA 18.5% faster)
- test_lora_models_loaded

Test runtime: ~0.6s (fast, suitable for CI/CD)

Signed-off-by: Yossi Ovadia <[email protected]>

---------

Signed-off-by: Yossi Ovadia <[email protected]>

Author: yossiovadia