Push ecosystem filters down to the parquet (fix per-document reload)

[tylere]

Why

EcosystemsGeoParquet.filter() loaded the entire national GeoParquet into memory (~11 GB for the Colombia map) then masked in-memory. Every crit_b document runs filter(code) in its own kernel, so a full render reloaded the whole national file ~85 times — turning the render into a multi-hour job (and it's the remaining bottleneck after the AOO-grid cache fix).

What

• filter(): for an exact match on a not-yet-loaded parquet, read only the matching rows via pyarrow predicate pushdown (filters=[(col, "==", value)]) instead of load() + mask. Regex and already-loaded data keep the in-memory base behavior. Result columns are unchanged (behavior-identical rows).
• _read() / _load(): read gs:// through pyarrow's native GCS backend (row-group pruning via range reads, no gcsfs), consistent with _remote_file_exists and the cache read/write path. http(s)/s3/az still use fsspec.

Pushdown only prunes when the parquet is clustered by ecosystem (small row groups). Paired with an ecosystem-sorted gs:// copy of the data, filtering one ecosystem drops from ~11 GB / 60 s to 0.39 GB / 3 s (measured end-to-end through the Ecosystems API).

Verification

• New TestEcosystemsGeoParquet tests: pushdown avoids full-load (_cached stays None), pushdown result equals full-load + mask (byte-identical rows), regex falls back to in-memory. Written test-first (the no-full-load test fails on the old code).
• Full test_ecosystems.py + test_aoo.py pass (except the pre-existing TestEcosystemsToRaster libgdal-dylib env failures, unrelated).

🤖 Generated with Claude Code