Add vector support for non-Postgres database loaders

[jamessewell]

Context

The CLI loader now validates vector(n) schema columns and currently allows them only for ParadeDB/PostgreSQL. Other backends can load scalar CSV/Parquet data, but vector columns are rejected early because each target database needs backend-specific schema mapping, ingest conversion, and query/index behavior.

Scope

• Define the supported vector representation per backend, including ClickHouse, Elasticsearch, OpenSearch, and MongoDB.
• Add backend capability metadata instead of hard-coding vector support to Postgres-style backends.
• Implement schema creation and ingest conversion for supported vector-capable databases.
• Add CSV and Parquet vector fixtures for each backend that supports vectors.
• Update the loader support matrix and dataset docs once support is added.

Acceptance criteria

• vector(n) schemas either load successfully for a backend with documented mapping, or fail with a targeted unsupported-backend error.
• Vector dimension validation is preserved across CSV and Parquet inputs.
• Each supported backend has unit or integration coverage for vector ingest.
• Docs list the exact vector type/mapping used by each database.