fix(python): preserve PQ num_bits in model training

[ddupg]

Background

While building vector indexes through Ray-based distributed indexing, we found that the PQ num_bits option could be ignored by the global PQ training path. The codebook is parameterized by this value: for example, 4-bit PQ should train 16 centroids per sub-vector, while 8-bit PQ trains 256.

Before this change, Python/PyO3 PQ training always used 8 bits internally. That can produce a PQ codebook whose bit width does not match the later index build settings, which can make pre-trained PQ artifacts inconsistent with segment construction and hurt quantization/index quality.

Summary

• expose num_bits on Python PQ training helpers and keep the default at 8 bits
• pass num_bits through the PyO3 PQ training path instead of hard-coding 8
• persist num_bits in saved PqModel metadata and reuse it when building from pre-trained PQ models

Tests

• make build PYTHON=3.12
• uv run --frozen pytest python/tests/test_indices.py::test_gen_pq python/tests/test_indices.py::test_indices_builder_multivector_distributed_dimensions
• uv run --frozen ruff check python/tests/test_indices.py
• uv run --frozen ruff format --check python/tests/test_indices.py