[C API] Add memory estimation functions for Vamana index build and search

[klmckeig]

Context

We need C API functions to estimate memory requirements before building a Vamana index. Currently we have no way to gate the build on available memory — we can only fail after the fact, after the user has already waited through a full table scan.

Use Case

Our goal is a pre-flight check in vamanabuild.c:

1. Estimate peak memory required given vector count, dimensions, and index parameters
2. Check whether maintenance_work_mem (or system memory) is sufficient
3. Build if yes; abort early with a clear error if no

Requested API

/**
 * Estimate peak memory (bytes) needed to build a Vamana index.
 *
 * @param builder       Configured builder handle (algorithm + storage set)
 * @param num_vectors   Number of vectors to be indexed
 * @param dimensions    Vector dimensionality
 * @param out_err       Error output handle
 * @return              Estimated peak memory in bytes, or 0 on error
 */
SVS_API size_t svs_index_estimate_build_memory(
    svs_index_builder_h builder,
    size_t num_vectors,
    size_t dimensions,
    svs_error_h out_err
);

/**
 * Estimate steady-state memory (bytes) of the built index (for search).
 *
 * @param builder       Configured builder handle (algorithm + storage set)
 * @param num_vectors   Number of vectors to be indexed
 * @param dimensions    Vector dimensionality
 * @param out_err       Error output handle
 * @return              Estimated in-memory index size in bytes, or 0 on error
 */
SVS_API size_t svs_index_estimate_search_memory(
    svs_index_builder_h builder,
    size_t num_vectors,
    size_t dimensions,
    svs_error_h out_err
);

Taking svs_index_builder_h as input means the estimates automatically reflect the configured storage type (float32/16, LVQ, LeanVec, SQ) and algorithm parameters (graph degree, window size, etc.) without us duplicating that logic in the extension.

Memory Components

Build-time peak (svs_index_estimate_build_memory):

• Graph adjacency list: num_vectors × (graph_max_degree + 1) × sizeof(Idx)
• Raw vector buffer: num_vectors × dimensions × sizeof(float) (needed before quantization)
• Temporary candidate pools / search history (scales with window_size, use_full_search_history, thread count)

Steady-state / search (svs_index_estimate_search_memory):

• Graph adjacency list (same formula)
• Stored vectors after compression (depends on storage type)
• Metadata (entry points, TID maps, etc.)

Upper-bound estimates are acceptable — we just need something accurate enough to make a go/no-go decision.

@mihaic @rfsaliev , please review the issue when you get a chance