fix: Add paginated merge and load-vocab-source command

[nicoloesch]

Summary

Fixes #12

• Adds paginated merge operations to orm-loader so large staging-to-target merges commit in bounded batches rather than one transaction.
• Adds a new load-vocab-source command to omop-alchemy with bulk mode, progress feedback, and crash-resilient retry.

orm-loader: paginated merge via _rownum

Staging tables now get a _rownum BIGINT GENERATED ALWAYS AS IDENTITY column at creation time. merge_insert, merge_replace, and merge_upsert all accept a merge_batch_size parameter (default 1 M rows). For tables larger than one batch, a _rownum index is built on the staging table and rows are processed in range-keyed batches, each committed independently. This bounds WAL accumulation to one batch per transaction instead of the full table. Small tables (below merge_batch_size) fall through to the original single-statement path.

The COPY statement was updated to include an explicit column list so the identity column is excluded from input.

omop-alchemy: load-vocab-source command

New cli_vocab.py implementing a load-vocab-source command with:

• --bulk-mode: disables FK triggers and drops indexes before loading, then rebuilds after. Substantially faster than per-table management for a full vocabulary reload.
• --merge-strategy: replace, upsert, or insert_if_empty.
• --merge-batch-size: passed through to the orm-loader paginated merge.
• Progress bar with per-phase descriptions including the post-load index rebuild phase, which can take 15+ minutes on concept_ancestor.
• Crash-resilient retry: if a retryable connection error occurs mid-merge and the strategy is insert_if_empty, the partially loaded table is truncated before retrying. Safe because FK triggers are disabled via ALTER TABLE ... DISABLE TRIGGER ALL and that state persists across crash and recovery.