Add freeze() to drop per-shard codes/residuals for read-only indexes

[raphaelsty]

Summary

• Adds FastPlaid.freeze() which deletes the per-shard {i}.codes.npy / {i}.residuals.npy files and sets "frozen": true in metadata.json. The merged file remains and becomes the sole storage, cutting on-disk usage roughly in half.
• Adds FastPlaid.unfreeze() as the inverse: slices merged_codes.npy / merged_residuals.npy back into per-shard files using doclens.{i}.json, then drops the frozen key. The roundtrip is byte-identical — every file matches the original.
• Teaches the loader to skip the chunk scan when an index is frozen and mmap the merged file directly (slightly faster reload, less peak RAM).
• update() and delete() on a frozen index raise RuntimeError — both paths operate per shard and can't run without re-sharding (call unfreeze() first).

Motivation

After indexing, FastPlaid keeps two copies of the same compressed data on disk:

• per-shard {i}.codes.npy / {i}.residuals.npy (used by update.rs / delete.rs)
• a merged merged_codes.npy / merged_residuals.npy (used at search time via StridedTensor)

For workloads that won't mutate the index again, the shards are dead weight. freeze() makes that explicit and reclaims the space without any Rust changes. unfreeze() is there as an escape hatch if you change your mind.

Byte-identical roundtrip

unfreeze() reconstructs shards via a small helper that matches tch::Tensor::write_npy byte-for-byte:

• Rust-style shape formatting (R,C,) (trailing comma, no spaces) instead of np.save's (R, C).
• Explicit < byte order even for 1-byte types (np.save writes |).
• Header padded so the prologue (10 + header) is a multiple of 16, matching tch's alignment.

metadata.json is written with indent=2 to match serde_json::to_writer_pretty. The frozen key is fully removed on unfreeze() so the file matches its pre-freeze bytes exactly.

What's untouched

• Indexing path (create.rs) and the chunk-then-merge flow are unchanged. Peak RAM during create is preserved.
• Search hot path is identical — same merged mmap, same StridedTensor.
• Non-frozen indexes behave exactly as before.

Test plan

• test_freeze_removes_shards_and_search_still_works — search results identical before and after freeze(), shard files gone, merged files retained.
• test_freeze_persists_across_reload — a frozen index reopens correctly in a new FastPlaid instance and returns the same results.
• test_freeze_update_raises — update() on a frozen index raises RuntimeError.
• test_freeze_delete_raises — delete() on a frozen index raises RuntimeError.
• test_freeze_idempotent — calling freeze() twice is a no-op.
• test_freeze_unfreeze_roundtrip_byte_identical — create a multi-shard index, copy the directory, freeze+unfreeze the original, assert every file is byte-identical to the snapshot via filecmp.cmp(shallow=False).
• test_unfreeze_restores_update_capability — after unfreeze(), update() works again.
• test_unfreeze_idempotent — calling unfreeze() on a non-frozen index is a no-op.
• Existing TestBasicCreateAndSearch and TestUpdate suites still pass.