[codex] Preserve provider usage metadata

[Tanisha-Katara]

Summary

• add optional provider usage metadata to ModelResponse
• preserve that metadata through ModelResponse slicing and cache serialization
• populate it from Hugging Face inference-provider chat completion usage fields

Why

Inference providers can return billing/token usage details such as prompt tokens, completion tokens, and nested cache counters. LightEval currently keeps token ids for evaluation, but drops provider-reported usage metadata from inference-provider responses. Preserving the raw usage payload makes saved details and cached samples more useful for downstream cost and cache analysis without adding pricing logic or a dependency on any external cost tool.

Tests

• .venv/bin/python -m pytest tests/unit/models/test_model_output.py tests/unit/models/endpoints/test_inference_providers_model.py -q
• .venv/bin/python -m compileall -q src/lighteval/models/model_output.py src/lighteval/models/endpoints/inference_providers_model.py tests/unit/models/test_model_output.py tests/unit/models/endpoints/test_inference_providers_model.py
• .venv/bin/ruff check src/lighteval/models/model_output.py src/lighteval/models/endpoints/inference_providers_model.py tests/unit/models/test_model_output.py tests/unit/models/endpoints/test_inference_providers_model.py
• .venv/bin/ruff format --check src/lighteval/models/model_output.py src/lighteval/models/endpoints/inference_providers_model.py tests/unit/models/test_model_output.py tests/unit/models/endpoints/test_inference_providers_model.py
• git diff --check