feat(test-emitter): Add evaluation performance testing framework

[zhirafovod]

Adds a package with perf test script and a test emitter, which samples and concurrently enqueues invocations and help to validate concurrent performance of the evaluator. The evaluator was run with a tiny 1.2B local model and mostly focused on correct metrics for every sampled invocation. Takes about 250s for 120 spans with 50% sampling on M3 mac.

Core changes:

• util/opentelemetry-util-genai-emitters-test new package with perf testing program and with the test emitter. Make sure to reinstall util/opentelemetry-util-genai-emitters* for proper registration
• util/opentelemetry-util-genai-evals/src/opentelemetry/util/genai/evals/manager.py - fixes/additions of helper functions to eval manager
• docs/feat-evals-perf.md - feature design and implementation details.
• dumps captured invocations and evaluations results, as well as detailed failures list (both bad evals or failures of evals) in /var/tmp by default

Features:

• Test Emitter captures all telemetry (spans, metrics, events, evaluation results) in memory
• CLI tool for testing evaluation framework performance and validating metrics
• 120 test samples across 6 categories (neutral, subtle_bias, subtle_toxicity, hallucination, irrelevant, negative_sentiment)
• Trace-based sampling with configurable sample rate
• Concurrent evaluation mode with configurable workers
• Threshold-based validation with score deviation metrics (MAE, RMSE)
• Idle timeout (60s) to prevent hanging on stalled evaluations
• JSON export of results and failures for debugging

CLI Options:

• --samples N: Number of test samples to use (default: 20)
• --concurrent: Enable concurrent evaluation mode
• --workers N: Number of concurrent workers (default: 4)
• --sample-rate N: Evaluate every Nth trace (default: 2 = 50%)- --sample-rate N: Evaluate every Nth trace (default: 2 = 50%)- --sample-rate N: Evaluate every Nth trace (default: 2 = 50%)- --sample-rate N: Evaluate every Nth trace (def_CONCURRENT: Enable- --sample-rate N: Evaluate every Nth trace (default: 2 = 50%)- --sample-rate N: Evaluate eSE_URL: Custom LLM endpoint
• DEEPEVAL_LLM_BASE_URL: Model URL for DeepEval
• DEEPEVAL_LLM_MODEL: Model name for DeepEval