Fix cosine normalization for small-norm float32 vectors

[shoaib0657]

Addresses #685.

For a component around 1e-25, squaring it in float underflows to zero. The current normalization then produces cosine distances around -1e10 and +1e10 instead of 0 and 2 for identical and opposite vectors.

This fixes normalization in both Index and BFIndex. It keeps float arithmetic for ordinary vectors and recomputes tiny or overflowing norms in double. Zero vectors keep their existing behavior, and normalized vectors are still stored as float32.

Tests

Added regression tests for tiny and large vectors, different stored/query scales, float boundaries, and zero vectors. The tiny-vector test reproduces the original bug.

Local checks on Ubuntu WSL2 with GCC 13.3, Python 3.12.3, and NumPy 2.5.3:

• make test: 25 tests passed.
• make examples: all five examples passed.
• git diff --check: passed.

Performance

Measured BFIndex.add_items on 50,000 random float32 vectors of dimension 128: seed 685, one thread, one warmup, and 15 timed rounds per version. Hardware: Intel Core i5-11320H under WSL2. Both builds used -O3/C++11 without -march=native; the order of the versions was randomized each round.

Version	Median time
Upstream	19.80 ms
This PR	19.59 ms

Insertion time is close to upstream on these ordinary inputs. The timing includes normalization and copying. HNSW build/query performance and workloads dominated by the double fallback were not measured.

Previously saved indexes affected by this bug need to be rebuilt from the original vectors.