Guidance 0.2.4

Better sampling, better metrics, llama-cpp-python fixes (update to latest please!), uncountable visualization fixes.

Added

• Allow changing sampling_params (top_p/top_k, min_p, repetition_penalty) on the fly Model.with_sampling_params(...)
• Add top_k tokens back into vis after temporary removal in previous refactor

Removed

• Model.token_count removed in favor of (currently private) Model._get_usage().output_tokens

Changed

• Bookkeeping of metrics such as input_tokens, output_tokens, ff_tokens, token_savings, avg_latency_ms have been added to State and are now accessible via (private for now) Model._get_usage(). This replaces bookkeeping that was previously attached to Engine instances.
• Factory functions create_azure_openai_model() and create_azure_openai_model() for accessing models hosted in AzureAI

Fixed

• Intermittent double widget render fixed.
• Widget doesn't always complete running, fixed.
• Widget backtracking bug fixed
• Widget now always show both inputs and outputs, sometimes would fail.
• TraceHandler forests stripped of extra trace nodes, sometimes caused render glitches.
• Widget latency displays now render.
• Widget early race condition resolved (sometimes widget is ready after backend is firing messages)
• Various linting and build improvements
• Tokens generated with OpenAI now correctly tagged as generated for vis
• Fix compatability with llama-cpp-python 0.3.12, bump dependency from 0.3.9 to 0.3.12 (first contrib: @jovemexausto)