### What happened?

`litellm/llms/databricks/cost_calculator.py` computes cost by hand:

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

This multiplies every prompt token by the flat `input_cost_per_token` and ignores the rest of the `Usage` block: `prompt_tokens_details.cached_tokens`, `cache_creation_tokens`, audio tokens, and reasoning tokens. It is the same shape as the Fireworks bug fixed in #33714, where the provider had a bespoke `cost_per_token` instead of delegating to `generic_cost_per_token`.

DeepSeek, xAI, Perplexity, Fireworks and other OpenAI-compatible providers delegate to `generic_cost_per_token`, which handles cache-read / cache-creation / audio / reasoning against the fields already on `model_info`. Databricks is an outlier.

### Impact

Latent today: no `litellm_provider: "databricks"` entry in `model_prices_and_context_window.json` currently publishes `cache_read_input_token_cost` or audio rates, so there is no active mis-billing. But Databricks serves models (e.g. Claude, Llama, DBRX) that can report cached tokens, so the moment …