fix: validate batch_size and empty-query benchmarks

[HarshRajSinghania]

Summary

Guard two unhandled boundary cases in ingestion and the benchmark harness.

Motivation

• ingest_documents(batch_size=0) dies with raw range() error #22: ingest_documents(..., batch_size=0) (via batched) raised ValueError: range() arg 3 must not be zero.
• run_benchmark ZeroDivisionErrors on a zero-query dataset #23: run_benchmark on a dataset with zero queries raised ZeroDivisionError from sum(r5) / len(r5).

Both issues are open, unassigned, labeled help wanted / good first issue, with explicit acceptance criteria. No open PRs existed for them.

Implementation

• batched() now raises ValueError("batch_size must be >= 1") when size < 1.
• run_benchmark treats empty per-mode recall lists as 0.0 instead of dividing by zero. Mode results are still emitted with n_queries=0 and qps=0.0. Index recall for non-dense / empty cases remains NaN as before.

Testing

Ran locally:

PYTHONPATH=src python -m pytest tests -q

Result: 34 passed.

New coverage:

• tests/test_ingest.py: batched([1,2,3], 0) and batch_size=-1 raise the specified error; ingest_documents(..., batch_size=0) does the same.
• tests/test_benchmark.py::test_benchmark_zero_queries_returns_empty_modes: empty query set returns modes with zeroed metrics.

Fixes #22
Fixes #23

[royalpinto007]

Thanks! Merging #26 for this one since it also guards ingest_documents directly (fail-fast) with lighter mock-based tests. Appreciate the work.