Extend OpenAITelemetryPlugin: Add tracking & pricing for cached tokens

[waldekmastykarz]

Feature Request

Extend the OpenAITelemetryPlugin to support token tracking and pricing information for cached tokens.

---

Background
Currently, the OpenAITelemetryPlugin tracks and calculates token usage and associated pricing. However, many OpenAI APIs (especially when caching is present) price cached tokens differently than non-cached tokens.

Requested Enhancement

• Update the plugin to distinguish between regular and cached tokens.
• Track cached tokens separately and calculate/report cost accurately based on their distinct rates.
• Factor in OpenAI pricing models that treat cached tokens differently, and expose the split and totals in pricing information/output.
• Update relevant reporting and metrics methods to show cached token counts and costs distinctly.

Impacted Code References:

• PricesData.cs / ModelPrices – only has Input and Output pricing, no CachedInput field
• PricesData.cs / CalculateCost() – does not account for cached tokens
• OpenAITelemetryPlugin.cs / RecordUsageMetrics() – passes entire PromptTokens value for cost
• OpenAITelemetryPlugin.cs / GetReportModelUsageInfo() – reads and reports cached tokens, but doesn't factor into cost logic
• LanguageModelPricingLoader.cs – supports prices only for input and output fields. Add support for cached_input if available

Benefits

• Improved reporting and transparency for users with cache-aware pricing.
• More accurate cost/control for workloads with cache/usage mix.
• Better insights in telemetry and exported reports for complex scenarios.