Feature: introduce incremental snapshotting

[desertfury]

Motivation

Persisting a large HNSW index with save is a stop-the-world operation:
the index serves no reads or writes for the entire flush. This PR adds an
interruptible, resumable serialization path and a global_rebuild_gt
adapter that periodically reconstructs and persists the index without
that pause - and without doubling memory.

Resumable serialization

index.hpp / index_dense.hpp gain a save_to_stream(output, state&, budget)
overload alongside the existing blocking one:

• index_serialized_state_t / index_dense_serialized_state_t are
  continuation cursors (stage + position + frozen counts).
• Each call writes at most budget nodes/vectors, then returns; pass the
  same cursor back to resume. The byte layout is identical to the blocking
  save_to_stream, so the file loads with the regular load.

global_rebuild.hpp - the orchestrator

global_rebuild_gt wraps a live "primary" index and drives a rebuild in
small budgeted steps, so the index keeps serving traffic throughout:

1. migrate - re-insert the primary's key-set (as of begin()) into a
   fresh "shadow" peer, reconstructing the HNSW graph from scratch;
2. save - stream the now-frozen shadow to disk via the resumable
   save_to_stream;
3. done - close the file, release the shadow, replay deferred removals.

Concurrent-mutation routing: add always goes to the primary; remove
during a rebuild is tombstoned and replayed at completion, so the on-disk
snapshot equals the begin() generation exactly.

The idea behind that adapter is described at Chapter 5 of The Design of Dynamic Data Structures book (https://link.springer.com/book/10.1007/BFb0014927)

Memory: no doubling

The shadow needs its own graph, but not a second copy of the vectors.
It is built with add(..., copy_vector = false), so every shadow node
references the primary's stored vector bytes (new index_dense_gt:: vector_data accessor) instead of duplicating them. Peak rebuild overhead
drops from vectors + graph (~2x) to one graph, and is released at
completion.

Testing

• test_global_rebuild - drives a non-blocking rebuild on a clustered
  12k x 256d dataset with interleaved add/remove, reloads the persisted
  file, and asserts: recall is preserved (1.000 -> 1.000), the shadow
  duplicates zero vectors (memory_stats().vectors_allocated == 0), and
  the process RSS sampled across the rebuild grows ~10% - never doubling.

[desertfury]

I also ran global_rebuild_gt through the insert-heavy tests to confirm the adapter doesn't fall over under a variety of workloads. And wired the adapter into test_collection across the whole matrix (scalars, connectivity, dimensions, collection sizes). Didnt commit that one