Preserve lazy shard loading by decoupling centroid routing from shard body loads

[rmax]

Index routing currently calls load_shard() for every shard just to read centroids, so the first query deserializes the full index into RAM before nprobe selection. This defeats the documented lazy-loading design, inflates cold-start latency, and makes memory usage scale with full index size instead of the probed working set.

Desired direction:

• separate centroid routing metadata from full shard bodies
• choose probe shards without loading every shard payload
• keep full shard bodies lazily loaded only for the selected probe set
• version any manifest or binary format changes explicitly
• update docs when formats change

Acceptance criteria:

• IndexSearcher can route without load_shard() on every shard
• first query only deserializes probed shard bodies
• tests verify non-probed shards are not loaded during routing
• docs updated for any schema/artifact changes
• benchmarking or instrumentation can confirm cold-start improvement