Add storing of prediction data for later training and debugging

We have two pending issues for Langfuse:

While it seems to be great for general prompt tunning (looking into prompt-completion pairs, costs, changing the prompt and seeing the result, annotating, etc.) it seems to be suboptimal for storage of arbitrary prediction-related data.

But before using anything else, check out if we can:

store arbitrary large and complex data in langfuse in reasonable manner
- for example, prediction market dump, scraped websites, etc.
if we can update the data with resolution, after it's known
- if we can index them and query them effectively
- for example, if we would like all the data between dates X and Y for the agent A where he was wrong, do we need to query everything and filter manually (would be problem in the future) or filter for them as with SQL in Postgres?

If all seems okay with Langfuse, let's use that, if not, Postgres is our go-to choice atm. We need:

Dumped prediction market data at the time of prediction
Queries to search and their results (including scraped data)
- Will be partially done for Tavily by https://github.com/gnosis/prediction-market-agent-tooling/issues/336
Any other data that are obtained and used for prediction (for example, correlated markets in think thoroughly)
Also, the prompt-completion pairs (as currently available in Langfuse, but in a way that we can easily query → Postgres)
Some way how to do this universally across agents, especially in multi-agent ones like CrewAI this will be tricky.

gnosis / prediction-market-agent

Add storing of prediction data for later training and debugging #374