refactor: extract stores and conformance tests into zarr-storage

[d-v-b]

🤖 AI text below 🤖

Extract the existing storage layer into zarr-storage, with its current APIs in zarr_storage.legacy. The package includes the Store ABC and byte request types; MemoryStore, ManagedMemoryStore, GpuMemoryStore, LocalStore, ZipStore, FsspecStore, and ObjectStore; WrapperStore, LoggingStore, and LatencyStore; StorePath and store utilities; and experimental CacheStore.

Context for this extraction:

• zarr-python#4217 — fully separate sync and async internal APIs explicitly proposes zarr-storage and zarr-codec subpackages that first reproduce the existing APIs, followed by new interfaces and a bridged migration. It also connects this work to modular array backends and zarrista.
• zarr-python#4317 — how to integrate subpackages proposes testing subpackages against Zarr first, adopting compatible runtime replacements next, and introducing new features before planning major API changes.
• zarr-python#3913 — upstream basic metadata classes and #4268 — subpackage for chunk key encodings give the broader packaging rationale: narrowly scoped components that other Python projects can reuse without depending on the whole array implementation.

These are design and migration references; this draft is an initial extraction and does not complete the proposed API redesign or remove the remaining dependency on zarr.

Include the complete store test suite and cache-store tests, redirected to the extracted implementations. Distribute StoreTests, state machines, strategies, buffer test utilities, and assertion helpers under zarr_storage.testing, with an optional testing extra. LatencyStore is separated from the pytest-dependent module so it can be used without pytest.

This remains an initial extraction: buffers, configuration, concurrency, and some metadata/sync helpers still come from zarr. The package must not become a Zarr dependency until this reverse dependency is removed. Zarr runtime imports remain unchanged, and the extracted classes have distinct identities. Array/group integration tests use a documented, test-only re-export fixture to exercise real Zarr arrays with the extracted stores; normal imports do not alter Zarr's bindings. No new storage contract or deprecation is introduced.

Validation:

• Built wheel, Python 3.12 with optional backends: 1,030 passed, 209 skipped, including mock-S3 tests.
• Built wheel, Python 3.14 minimal: 847 passed, 193 skipped.
• Source, Python 3.12 minimal: 847 passed, 193 skipped.
• Opt-in stateful tests: 3 passed, 3 upstream capability skips.
• Wheel and source distribution builds pass. Signature parity, normal-import isolation, and pytest-free LatencyStore imports are covered.

CI covers source and built-wheel tests on Python 3.12/3.14, with a Python 3.12 optional-backend job and an explicit stateful test step. GPU tests skip without appropriate hardware, and existing unsupported-backend-operation skips are preserved.