fix: PGVector.add_embeddings silently truncates on length mismatch

[Humphrey (HumphreySun98)]

Description

`PGVector.add_embeddings` (and `aadd_embeddings`) build the SQL insert payload via:

```python
data = [
{...}
for text, metadata, embedding, id in zip(texts, metadatas, embeddings, ids_)
]
return ids_
```

Plain `zip` truncates to the shortest argument, then the method returns the full `ids_` list regardless. So when an upstream embedder bug returns fewer embeddings than texts — e.g. langchain-ai/langchain-google#1704's `gemini-embedding-2` regression — the caller sees N IDs but only M rows actually land in Postgres, with no exception. This silent data loss is exactly the symptom reported in #300.

This PR validates the input lengths up-front via a new `_validate_lengths_match` helper that raises a clear `ValueError` before any DB session is opened, and adds `strict=True` to the trailing `zip` calls as a backstop. Both sync and async paths are covered.

Relevant issues

Fixes #300

Changes

• `langchain_postgres/vectorstores.py`:
  • Add module-level `_validate_lengths_match(texts, embeddings, metadatas, ids)` raising `ValueError` with the actual lengths in the message.
  • Call it at the top of `add_embeddings` and `aadd_embeddings` (after the sync/async early-init blocks, before any DB session opens).
  • Add `strict=True` to the trailing `zip` in both methods so any future internal mismatch turns into an exception instead of a silent truncation.
• `tests/unit_tests/v1/test_pgvector_length_validation.py`:
  • Five focused tests that bypass `PGVector.init` (no Postgres needed):
    • too few embeddings → `ValueError`
    • mismatched metadatas length → `ValueError`
    • mismatched ids length → `ValueError`
    • matching lengths → validation passes through (test asserts it reaches the DB-session line)
    • async `aadd_embeddings` analog

Testing

```
$ python -m pytest tests/unit_tests/v1/test_pgvector_length_validation.py -v
5 passed in 0.64s
```

Reverting only the `_validate_lengths_match` call and `strict=True` flag while keeping the new tests makes the mismatch tests fail back to silent truncation, confirming the regression coverage pins the buggy behavior. `ruff check` and `ruff format --check` pass.

Note

The same anti-pattern exists in `langchain_postgres/v2/async_vectorstore.py:297` (`for id, content, embedding, metadata in zip(ids, texts, embeddings, metadatas):`). v1 is the path explicitly called out in #300 and the path users hit through the deprecated `PGVector` import; happy to fix the v2 analog in a follow-up if you'd like it bundled differently.

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.