fix: defer PostgresChatMessageHistory import to keep psycopg lazy

[Humphrey (HumphreySun98)]

Description

`langchain_postgres/init.py` eagerly imports `chat_message_histories.PostgresChatMessageHistory`, which top-level imports `psycopg` (and `psycopg.sql`). So `from langchain_postgres import PGEngine` — or anything else from the driver-agnostic v2 API — transitively loads the LGPL-3.0-licensed `psycopg`, even for downstreams that only use `PGEngine` / `AsyncPGVectorStore` / `Column` on an Apache-2.0 driver like `asyncpg`. As reported in #296, this currently forces those users into a fork or post-install patching step.

This PR moves `PostgresChatMessageHistory` behind a module-level `getattr` so it's only imported when actually accessed. `PGVector` (which doesn't directly depend on `psycopg`) is unchanged.

This is the minimal of the three options the issue suggested — "lazy-load the v1 surfaces". The other two (declaring `psycopg` as an optional extra, or splitting the package) are larger and require maintainer direction; this change is a low-risk first step that solves the reported import-side-effect without changing public API.

Relevant issues

Fixes #296

Changes

• `langchain_postgres/init.py`: replace the eager `from langchain_postgres.chat_message_histories import PostgresChatMessageHistory` with a module-level `getattr` that performs the import on first access. `all` is unchanged so `PostgresChatMessageHistory` remains a documented public symbol. A `TYPE_CHECKING` import keeps static analyzers happy.
• `tests/unit_tests/test_lazy_chat_message_histories.py`: four tests
  • `PostgresChatMessageHistory` accessed via the package attribute path resolves to the same class as the direct submodule import (no backward-compat regression);
  • unknown attributes raise `AttributeError` (the `getattr` fallback works);
  • a fresh subprocess that imports only `PGEngine` / `PGVectorStore` has no `psycopg` entry in `sys.modules`;
  • the same subprocess does pick up `psycopg` once the lazy attribute is touched (so the v1 chat-history API keeps working when needed).

Testing

```
$ python -m pytest tests/unit_tests/test_lazy_chat_message_histories.py -v
4 passed in 2.82s
```

Reverting the `init.py` change while keeping the new tests makes the subprocess test `test_psycopg_not_imported_when_only_v2_symbols_used` fail (`psycopg` shows up in `sys.modules`), confirming the regression coverage pins the buggy behavior. `ruff check` and `ruff format --check` pass.

Note

Happy to follow up with the optional-extra approach (option 2 from #296) or a docs note if you'd prefer either alongside this change — let me know.

Disclaimer: this PR was prepared with the assistance of an AI agent (Claude Code). All code and test changes were reviewed by the author before submission.