perf(coco): right-size index reserve and switch to FxHash

[Borda]

This pull request significantly optimizes the COCO::create_index method and related data structures for handling COCO datasets, focusing on performance and memory efficiency. The main changes include switching to a faster hash map implementation, optimizing how indices are built (especially for large datasets), and adding comprehensive tests to ensure correctness and maintain important invariants.

Performance and Data Structure Optimizations:

• Switched all internal index maps in the COCO struct from the standard library's HashMap to rustc_hash::FxHashMap, which is much faster for integer and integer-pair keys. This change improves performance, especially on large datasets. [1] [2] [3] [4]
• Optimized img_cat_to_anns to reserve capacity based on the number of distinct (image, category) pairs instead of the total number of annotations, reducing memory waste for dense workloads.
• Changed the construction of cat_to_imgs to derive its entries from the set of unique (img, cat) pairs, rather than pushing once per annotation. This results in fewer insertions and guarantees deduplication.
• Ensured that cat_to_imgs is sorted and deduplicated after construction for deterministic iteration order and safety, even though the construction method now guarantees uniqueness.

Testing and Documentation:

• Added a thorough test (test_cat_to_imgs_derived_from_pair_keys) to verify that cat_to_imgs is built from unique (img, cat) pairs and that img_cat_to_anns preserves annotation order, maintaining compatibility with COCO evaluation semantics.

Changelog and Dependency Updates:

• Updated the changelog to document these optimizations and the switch to rustc-hash as a direct dependency. [1] [2]
• Added rustc-hash as a direct dependency in Cargo.toml.

---

• Reserve img_cat_to_anns for distinct (img, cat) pairs, not the annotation count (~400K vs ~1.5M on the RF-DETR workload).
• Derive cat_to_imgs from img_cat_to_anns's unique keys after the annotation loop instead of pushing once per annotation.
• Switch all six index maps (anns, imgs, cats, img_to_anns, cat_to_imgs, img_cat_to_anns) from std HashMap to rustc_hash::FxHashMap. All six are private; no public signature changes. Map iteration never reaches output (every iteration site feeds a sort), verified before the swap.
• Add rustc-hash as a direct dependency (already present transitively via numpy); MIT/Apache-2.0, passes deny.toml.
• No value changes: precision/recall/scores/stats bit-identical to main on 10 freshly-generated configs (the original baseline dumps' exact args for two configs were unrecoverable, so a new set was authored and dumped from a throwaway main worktree).
• loadRes 34% faster, COCOeval ctor 49% faster, end-to-end 20% faster on the RF-DETR-shaped harness workload (both phases call create_index).
• Add test_cat_to_imgs_derived_from_pair_keys, pinning cat_to_imgs's membership/dedup and get_ann_ids_for_img_cat's dataset-order contract; verified it fails on three injected defects (swapped pair fields, wrong source map, sorted values).