refactor(metadata-db): decouple physical tables from storage paths

[shiyasmohd]

Split physical_tables into two tables to enable data sharing across
datasets with the same spec:

• physical_tables: meta table with dataset identity and active revision pointer
• physical_table_revisions: storage paths and writer info

This allows multiple datasets to share the same physical storage while
maintaining independent activation states per (namespace, name, hash, table).

Changes:

• Add migration to split tables and preserve FK compatibility
• Update unique constraint to (dataset_namespace, dataset_name, manifest_hash, table_name)
• Modify register/mark_active/mark_inactive APIs to include dataset identifiers
• Update all queries to join tables appropriately
• Add index on active_revision_id for query performance

[LNSD]

As this PR is changing the way we manage physical tables, I think we should add/update a docs/features/*.md document to reflect the new "link/unlink" behaviour.