Fix/reverse search doc and validation

[ahmedkansulum]

Summary

While working with VectorStore I noticed that reverse_search's
docstring does not match its behaviour:

• The implementation simply joins a list of document IDs onto the existing
  vector store and returns the matching rows.
• The docstring, however, describes a cosine-similarity based search and
  mentions n_results, which is not used at all.

In addition, reverse_search accepts an optional ids argument but does
not validate it. If ids has the wrong length or contains incompatible
types, this can lead to confusing errors from Polars rather than a clear
ValueError.

This PR tidies up both of these points without changing the behaviour of
reverse_search.

---

Changes

• Update the VectorStore.reverse_search docstring to accurately describe
  the current behaviour:

  • Emphasise that the method looks up rows by document id and does not
    compute new embeddings or similarity scores.
  • Clarify the roles of query, ids, and the returned columns.
  • Document that n_results is currently unused but retained for public
    API compatibility.

• Add input validation for the optional ids argument:

  • Require ids to be:
    • a list,
    • containing only str or int values,
    • the same length as query, and
    • containing only unique values.
  • If these conditions are not met, raise a clear ValueError with a
    helpful message (mirroring the behaviour already used in
    VectorStore.search).

• Add a short inline comment noting that n_results is currently unused.

---

Rationale

• Keeps the public docstrings in sync with the actual implementation,
  reducing confusion for users of the public API.
• Makes the behaviour of reverse_search more robust and predictable
  when a caller provides custom ids.
• Aligns argument validation between search and reverse_search:
  both now enforce a consistent contract for ids.

---

Testing

Locally I:

make check-python