surtgis v0.10.3

Patch release that adds the benchmark suite for the extract-patches
subcommand. Companion-paper submission to Computers & Geosciences
("Compute-efficient preparation of training data for geospatial
foundation models in native Rust") cites this version as the
reproducibility anchor for §6 of the manuscript.

Added

• benchmarks/measure_memory_gfm_prep.sh — peak resident-set
  measurement via /usr/bin/time -v. Writes
  benchmarks/results/gfm_prep/memory.csv.
• benchmarks/verify_outputs_gfm_prep.sh — SHA-256 hashes plus
  element-wise diff statistics on the single-timestamp comparable
  workload. Writes output_verification.txt.
• benchmarks/run_gfm_prep_sweeps.sh — scaling sweeps over point
  count, patch size, and timestamp count. Forces LC_ALL=C for
  locale-independent printf. Writes
  benchmarks/results/gfm_prep/scaling.csv.
• benchmarks/run_gfm_prep_3way.sh — 3-way comparison harness
  (SurtGIS vs naive Python vs InstaGeo-style xarray+rioxarray
  proxy). Writes three_way.csv.
• benchmarks/bench_gfm_prep_instageo_style.py — InstaGeo-style
  chip-extraction proxy that reproduces the
  RasterDataPipeline.process_tile inner loop on local data
  without installing the CC-BY-NC-SA-licensed instageo package.
• benchmarks/measure_stac_latency.py — one-off STAC query
  latency probe against Element 84 Earth Search v1.

Notes

• All five new bench scripts and their committed result files
  (memory.csv, scaling.csv, three_way.csv, output_verification.txt)
  are the reproducibility anchors cited in the paper-gfm-prep
  submission. The paper's repository
  (github.com/franciscoparrao/paper-gfm-prep)
  references these files by path; the v0.10.3 tag ensures the
  cited paths resolve.
• No algorithm changes vs v0.10.2. Same library functionality;
  only the benchmark suite expands.