…ect_store

ZarrChunkReader now accepts file://, bare paths, s3://, gs:// / gcs://,
az:// / abfs:// / abfss://, and http:// / https://. The reader's
public API is pure-storage-in: hand it an
Arc<dyn AsyncReadableListableStorageTraits> and a group URI (for
chunk-anchor URI formatting) and it walks the chunk grid. Where the
storage came from — registry lookup, env-var-built builder, an
in-memory test fixture — is the caller's concern, not the reader's.

* `open_storage_from_uri(uri, store_override: Option<Arc<dyn ObjectStore>>)`
  is the helper for callers that don't already hold a credentialed
  store. `Some(store)` uses it directly (with PrefixStore for
  s3/gs/az; http(s) is rooted at the URL by its builder). `None`
  falls back to per-scheme *Builder::from_env().with_url(uri).build()
  — the same env-var credential discovery that read_format's
  orchestrated path gets via ensure_object_store_registered_with_options.
  file:// and bare paths always go through zarrs_filesystem's
  FilesystemStore wrapped in SyncToAsyncStorageAdapter, so the local
  backend shows up on the same async storage surface.
* ZarrChunkReader::try_new is async; the sync RecordBatchReader
  streaming surface is unchanged since next() is pure CPU.
* No dedicated tokio runtime; async tasks run on whatever runtime
  the caller is on (DataFusion's executor for the future SQL UDTF,
  an ad-hoc current-thread runtime block_on'd in PyZarrChunkReader::new
  for the Python FFI, #[tokio::test] in tests).
* PyZarrChunkReader::new builds storage via open_storage_from_uri
  (env-var credentials by default), then drives the async try_new
  through a local current-thread tokio runtime. No synthetic
  RuntimeEnv, no datafusion-execution dep on the FFI cdylib.
* Replaces Group::child_arrays (which opens every child up front and
  errors hard on any per-array metadata failure) with two purpose-
  built paths driven by the caller's arrays filter:
  - arrays: Some([...]) opens each by name with Array::async_open.
    No listing — usable against backends that can't list (plain HTTPS
    without WebDAV, S3-via-HttpStore).
  - arrays: None lists direct children with storage.list_dir, then
    Array::async_open each. Per-array open failures are logged at
    warn! and skipped, so a single malformed sibling (e.g. an
    xarray-style fixed-length-Unicode coord variable with a null
    fill_value that zarrs 0.23 can't open) no longer poisons the
    rest of the group.
* zarrs_object_store pinned to 0.5 in the workspace; 0.6 depends on
  object_store 0.13, semver-incompatible with DataFusion 52's
  object_store 0.12.x.
* Cloud smoke tests pass strictly against the public anonymous
  ITS_LIVE v2 ice-velocity datacubes (s3://its-live-data/...). Same
  bucket via s3:// (AmazonS3Builder) and via the virtual-hosted
  HTTPS URL (HttpStore), with an explicit M11/M12 filter to avoid
  listing on the HTTPS path.

PyZarrChunkReader::new built a fresh current-thread tokio runtime on every
open to bridge the async loader constructor into the sync Python FFI. Replace
it with a process-wide OnceLock<Runtime> so the package builds the runtime
once and reuses it. next() on the reader is pure CPU and never touches it.