Improve langfuse tracing of lesson insights

[lfryemason]

Improvements:

• Fix folder that langfuse prompts are retrieved from
• Add prompt name and version to trace to associate with the langfuse prompt
• Change from sending in the entire system prompt to adding the specific variables - this will allow us to create datasets and test different prompts on the trace
• Add lesson feedback traces
• Filter out everything except the expected end output from the langfuse response so we can also add intermediate steps to the AI response and log them to langfuse
• Use the development tagged prompts in development

Now working - see correctly associating prompt version

Adds correct tracing variables

Also adds input variables

Links

• Jira: https://codedotorg.atlassian.net/browse/TEACHING-258 & https://codedotorg.atlassian.net/browse/TEACHING-250

[etaderhold]

I think there's some ideal where Honeybadger entries are supposed to represent actionable bugs that need to be fixed in our code. There's of course a lot of noise there right now, and adding more entries for things like external API timeouts outside of our control might make it even noisier. Most of the usages of notify_and_raise do seem like encountering them would represent an actionable bug fix opportunity though.