### What happened

`litellm/llms/fireworks_ai/cost_calculator.py :: cost_per_token()` bills **100% of prompt tokens at the full input price**, ignoring cached tokens entirely:

```python prompt_cost: float = usage["prompt_tokens"] * model_info["input_cost_per_token"] ```

It never reads `usage.prompt_tokens_details.cached_tokens`, and it never applies `model_info["cache_read_input_token_cost"]` — even though the docstring claims the usage block contains "caching information" and the model map for Fireworks models (e.g. `fireworks_ai/accounts/fireworks/models/glm-5p2`) already defines `cache_read_input_token_cost`. The generic cost calculator right next to it (`litellm/litellm_core_utils/llm_cost_calc/utils.py :: generic_cost_per_token`) handles both the cached-token split and the cache-read price; the Fireworks-specific path silently does not.

Fireworks serverless **does** cache and **does** return `prompt_tokens_details.cached_tokens` in usage (verified against the live API in both non-streaming and streaming), so on long agentic sessions with 80–95% cache hit rates this overbills input by up to ~4×.

### Repro (LiteLLM 1.91.0)

```python from litellm.types.utils import Usage, Pr…