[Enhance]: Support the Flat index compute distances through turbo Quantizer

### Affected Component

Flat Searcher, Flat Builder, Flat Streamer

### Current Behavior

The `turbo` module already provides registered quantizers (`Fp32Quantizer`, `Fp16Quantizer`,
`Int8Quantizer`, `Int4Quantizer`, `PqInt8Quantizer`) with own SIMD distance kernels and APIs.

Every distance in the Flat path goes through `FlatDistanceMatrix` / `IndexMetric`
(`distance_matrix().distance<1>(...)`, `row_distance_`, `column_distance_`). Quantized data can
therefore only be searched by going through a Converter/Reformer pair and a matching `IndexMetric`,
which duplicates logic that the turbo quantizers already implement (and better).

### Desired Improvement

Plumb an optional `std::shared_ptr` through the framework and make the Flat
index delegate distance computation to it when present.

### Impact

- Turbo quantizers become usable from the real Flat build/search flow instead of tests only, so
int8/int4/fp16 indexes get the turbo SIMD kernels (AVX2 / AVX‑512 / AVX‑512‑VNNI with scalar
fallback) rather than a Converter + generic `IndexMetric` chain.
- The quantizer choice is persisted in the index meta, so a quantized index is self-describing and
the query side is reconstructed automatically at load time.
- Zero behavior change for existing indexes: without `quantizer_name` in the meta, every code path
falls back to the current metric/distance-matrix logic.