feat: upgrade query adapter algorithm

[lsorber]

Changes:

1. Only insert missing evals in _bench.py.
2. Only compute the query adapter if missing in _bench.py.
3. Add a 'RAGLite with query adapter' to the benchmark command.
4. Output the current document id as a tqdm postfix in insert_documents.
5. Parallelise extraction of triplets in _query_adapter.py.
6. Improve the query adapter by introducing per-query target weights and optimize these on a validation set of evals with L-BFGS.
7. Add tests to verify that the gradient formula is correct.

[Copilot]

Pull Request Overview

This PR upgrades the query adapter algorithm by refactoring its implementation for multi-threaded triplet extraction, weight optimization via L-BFGS, and adds gradient-based unit tests. It also extends typing for float tensors, refines the CLI bench command with better error handling and reranker support, and enhances progress bar outputs in document insertion.

• Refactored _query_adapter.py to introduce helper functions (_extract_triplets, _optimize_query_target, _compute_query_adapter_grad, etc.), multi-threading, and weight optimization with SciPy’s minimize.
• Added FloatTensor alias in _typing.py and a test_query_adapter_grad in tests/test_query_adapter.py to validate the gradient.
• Improved the CLI bench command (_cli.py, _bench.py), including a prescore step, optional reranker integration, and user-friendly import errors.
• Minor enhancement: show inserted document IDs in the progress bar during _insert.py.

Reviewed Changes

Copilot reviewed 6 out of 6 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
tests/test_query_adapter.py	Added gradient correctness test using scipy.check_grad and increased num_evals for stability
src/raglite/_typing.py	Introduced FloatTensor alias for 3-D float arrays
src/raglite/_query_adapter.py	Full refactor of query-adapter optimization logic, threading, and weight learning
src/raglite/_insert.py	Show document ID in insertion progress bar
src/raglite/_cli.py	Guard bench imports, expose reranker option
src/raglite/_bench.py	Added prescore hook and reranker support in evaluator

Comments suppressed due to low confidence (1)

src/raglite/_query_adapter.py:80

• [nitpick] The _compute_query_adapter helper now implements the core transform logic for both 'dot' and 'cosine' metrics but lacks direct unit tests. Consider adding tests that verify its output under known inputs.

def _compute_query_adapter(