perf(convert): intern decode-path dict keys

[Borda]

This pull request optimizes the deserialization of Python dicts into datasets by interning field-name keys, significantly reducing string allocations and improving performance. All macro and manual dict field lookups in Rust now use pyo3::intern!, ensuring each key string is created only once per process. The change is validated by a comprehensive new test that round-trips every known field of annotations, images, and categories, ensuring no key is missed or misnamed.

Performance and memory improvements:

• All macros and manual field lookups in crates/hotcoco-pyo3/src/convert.rs now use pyo3::intern! to intern field-name keys, preventing millions of temporary string allocations during dataset decoding and yielding an 18–20% speedup for common workloads. [1] [2] [3] [4] [5] [6]

Testing and validation:

• Added TestKnownKeysRoundTrip in crates/hotcoco-pyo3/tests/test_dropin_gaps.py, which round-trips every known annotation, image, and category field through COCO(dict), ensuring all keys are correctly interned and decoded, with failures pinned for both required and optional fields.

Documentation and changelog:

• Updated CHANGELOG.md to document the key interning optimization, its performance impact, and the new test coverage.

These changes collectively improve dataset decoding efficiency and reliability, especially for large annotation sets.

---

• opt!/req!/opt_with!/req_with! macros and the standalone get_item calls in py_to_annotation/py_to_segmentation/py_to_rle now fetch each key via pyo3::intern! instead of a bare &str, which allocated a fresh PyString on every call; macro $key tightened from :expr to :literal to match what intern! requires
• decoding an RF-DETR-shaped annotation list calls get_item once per field per detection, so this removes millions of throwaway PyString allocations for a fixed set of ~10 field names; loadRes(list of dicts) routes through the same py_to_annotation and benefits too
• COCO(dict) construction is 18-20% faster across three shapes (small dict, 1.5M-annotation RLE segmentation dict, bbox-only dict), thread-count-invariant since the decode loop is sequential
• new TestKnownKeysRoundTrip round-trips every known annotation/image/ category field through COCO(dict) and fails if an interned key literal drifts from its field, proven by injection on both a required-key typo (raises) and an optional-key typo (silently drops the field)
• 10/10 bbox + 4/4 segm bit-identical against a fresh main build; no numeric or structural change to the decoded dataset
• extract_extra's key-to-String allocation and linear known-key scan is not touched here; reprofiling found it costs roughly half of what remains of decode time after this change, left open