Configurable context_window and tokenizer

[imartinez]

Extract two key model parameters to settings.

• Tokenizer: used to calculate the available context window given a prompt in several LlamaIndex pipelines. The default tokenizer is the one used for gpt-3.5 but it is not aligned with, for example, Mistral model. Using huggingface as provider of the tokenizer information, added to setup download script.
• context_window: base model argument

[cognitivetech]

is there a way to make this setup script not download the model if I already have it?

running setup for embeddings \ tokenizer.. since I don't know how to add those manually but it also triggers superfluous model download.

alternatively it would be nice to have a documented way to manually download and configure embeddings \ tokenizer