refactor(rust/sedona-raster): make ViewEntries the interface for passing views

[james-willis]

The smell

ViewEntries::new(input.view().to_vec()).compose(&ViewEntries::new(view.to_vec()))?

Two heap allocations to read two slices. It looks like a workaround because it is one.

Why it was there

ViewEntries existed as the container type, but nothing actually passed it around. BandRef::view() returned &[ViewEntry], and every struct field carrying a view held a bare slice:

	before
BandRef::view()	&[ViewEntry]
StartBandArgs::view	Option<&[ViewEntry]>
WithViewArgs::view	&[ViewEntry]
BandOverrides::view	Option<&[ViewEntry]>
RasterLoadRequest::view	&[ViewEntry]
RasterLoadResult::view	Vec<ViewEntry>

So slices were the real currency and ViewEntries was a method bag. compose was the only operation demanding the owned type on both sides — which is why every caller had to allocate its way in.

Two supporting details:

• BandRefImpl already stored a ViewEntries and called .as_slice() on the way out. Returning the container removes a conversion rather than adding one.
• Both doc comments describing this API were uncompilable as written — they said source.view().compose(&next)?, but view() returned a slice, which has no compose. They described the API this PR builds.

Changes

• BandRef::view() -> &ViewEntries
• compose's next takes impl AsRef<[ViewEntry]>, so a ViewEntries or a bare slice both work — existing a.compose(&b) call sites are untouched
• The six fields above all carry ViewEntries

The motivating site is now:

let effective_view = match overrides.view {
    Some(v) => self.view().compose(v)?,
    None => self.view().clone(),
};

A duplicate identity check, silently diverged

Chasing the same thread turned up rs_ensure_loaded::view_is_identity, a private copy of ViewEntries::is_identity — which had drifted from it:

	empty view, source_shape = [2, 3]
view_is_identity	true (short-circuits on is_empty())
ViewEntries::is_identity	false (length mismatch)

An empty view is identity only for a 0-dimensional raster, which is what the method already implements — the local copy was wrong.

It was unreachable in production: RasterLoadResult::unresolved clones the request's view, the request's view is the band's view, and start_band rejects 0-dimensional bands outright. The only things exercising the divergence were loader tests passing an empty view alongside a 2-D or 3-D source_shape — a state that cannot occur. Those fixtures now pass a real identity view over their source shape, and the helper is deleted in favour of the method.

No behaviour change

Composition, validation, and persistence are untouched. The only semantic change is that the identity check now has one definition instead of two that disagreed.

sedona-raster            183 passed; 0 failed
sedona-raster-functions  256 passed; 0 failed
sedona-raster-zarr        66 passed; 0 failed

fmt clean, clippy clean (no unused imports), sedona-raster-gdal builds.

Note

#1158 is stacked on this branch — it converts BandOverrides to a trinary Override<T>, which composes cleanly with the type change here.

[jiayuasu]

LGTM!