feat: add SINDI inverted index with SIMD-accelerated window-based sparse vector search

[sparknack]

Summary

Add the SINDI (SIMD-accelerated Inverted Index) for sparse vector search, based on the SINDI algorithm. SINDI partitions posting lists into fixed-size windows, storing only 16-bit local offsets instead of full 32-bit doc IDs. This significantly reduces memory footprint and enables efficient SIMD-based scoring.

Key Features

• Window-based TAAT search: For each window, scores are scatter-accumulated into a dense buffer via SIMD (AVX-512/AVX2/SVE), then batch-inserted into a top-k heap
• Platform-specific SIMD: Separate compilation units for AVX-512, AVX2, and SVE with runtime dispatch

Reference

• SINDI paper: https://arxiv.org/abs/2509.08395