perf(ogc): faster page parsing and chunk combination

[thodson-usgs]

Summary

Two client-side performance improvements to the OGC page-parsing and chunk-combination paths, identified through profiling the parallel_chunks fan-out code.

Changes

1. Replace pd.json_normalize with pd.DataFrame in _get_resp_data (shaping.py)

   OGC API feature properties are flat dictionaries — json_normalize's recursive nesting support is unnecessary overhead. Direct pd.DataFrame construction is 2.4× faster per page (3.2 ms → 1.35 ms on 2500-feature pages).

2. Skip drop_duplicates when the plan has no filter axis (planning.py, chunking.py)

   List-axis chunks (the common parallel_chunks case — splitting on monitoring_location_id, parameter_code, etc.) produce non-overlapping partitions by construction. Only filter-axis chunks (splitting on CQL OR clauses) can overlap. _combine_chunk_frames now accepts a dedup flag and ChunkedCall._combine_raw passes dedup=self.plan.has_filter_axis — 2.7× faster combine (3.4 ms → 1.25 ms on 52k-row results) for the common case.

Impact

For a typical 32-chunk query fetching 32 pages (~52k rows):

• Page parse savings: ~1.85 ms × 32 pages = ~59 ms saved
• Combine savings: ~2.2 ms saved
• Total: ~61 ms less client-side overhead per query

These are additive to the network-latency savings parallel_chunks already provides.

Testing

All 574 tests pass. No behavioral change — list-axis chunks were never producing duplicates (the dedup was a no-op); filter-axis chunks still dedup as before.

[thodson-usgs]

Closing as stale.

The branch is CONFLICTING, and it still carries feat(ogc): add opt-in parallel chunk fan-out — which landed separately as #341 — so the PR duplicates already-merged history. The ogc/ package has since been split up in #346 and #348, and is being restructured again in #350/#351, so the files this patches no longer look like this.

The two perf findings are real and are not on main yet. Worth redoing against current main:

1. pd.json_normalize → pd.DataFrame in _get_resp_data (now dataretrieval/ogc/shaping.py:126). OGC feature properties are flat dicts, so the recursive-nesting path is pure overhead — measured 2.4× faster per page (3.2 ms → 1.35 ms on 2500-feature pages).
2. Skip drop_duplicates when the plan has no filter axis (now dataretrieval/ogc/combining.py:144). List-axis chunks partition by construction, so the dedup pass can't find anything.

Branch perf/chunk-client-side is retained if we want to cherry-pick the perf commit.