Better support for token count and tokenization depending on the model in use

[dluc]

Motivation and Context (Why the change? What's the scenario?)

Allow to use different models with different tokenization characteristics.

High level description (Approach, Design)

• New interfaces required for text and embedding generation.
• Handlers, TextChunker and SearchClient use the current model to check token count, to check partition size, to check summarization settings, etc.
• Text generation backends must provide a MaxToken (configurable) in order to check for correct prompt sizes
• Embedding generation backends must provide a MaxToken (configurable) in order to check for correct partitioning/chunking
• All LLMs must provide a tokenizer, in order to count the token flowing through. The system falls back to the default GPT3 tokenizer if none is provided
• Embedding and Text model details can be configured in appsettings.json or passed with the usual "With*" builder methods