Infer and return the structural schema of data sources

Call this once when a user connects a data source.

Returns the full Algenta structural schema: column roles, detected formulas, cardinality, and query_hints the LLM should use when calling POST /v1/query.

The schema is also embedded as Arrow metadata in the returned Parquet bytes (parquet_b64), which the LLM agent can store and reuse — no re-profiling on subsequent queries.

Flow:

1. User connects data source
2. Call POST /v1/sources/schema — ~1-2s, runs once
3. LLM reads schema JSON and query_hints
4. User asks questions → LLM produces structured intent → POST /v1/query (~200ms, cached)