Update cuml.accel benchmark results and presentation

docs/benchmarks/cuml-accel/README.md

@@ -0,0 +1,35 @@
+# Synchronizing the cuml.accel benchmark page
+
+The checked-in `benchmark-data.json` is the portable, non-published input for
+the Sphinx page. It is produced and schema-validated by `cumlbench-dash`; raw
+mlbench results stay outside this repository.
+
+From the repository root, synchronize an updated publication artifact with:
+
+```console
+python docs/benchmarks/generate_cuml_accel_benchmarks.py sync \
+  --data /path/to/benchmark-data.json
+```
+
+The sync command validates publication schema version 1, content-addressed
+source provenance, record source references, and the 147-record transport
+shape, including each case's backend-neutral declared estimator parameters. It
+copies the artifact unchanged, selects the widest available PCA `fit_transform`
+result for PCA's canonical large case, and renders the resulting 145-record RST
+page and SVG heatmaps.
+
+Render the page and heatmaps, or verify that they are current, with:
+
+```console
+python docs/benchmarks/generate_cuml_accel_benchmarks.py render
+python docs/benchmarks/generate_cuml_accel_benchmarks.py sync --check \
+  --data /path/to/benchmark-data.json
+python docs/benchmarks/generate_cuml_accel_benchmarks.py render --check
+```
+
+`sync --check` validates both the supplied and checked-in artifacts, verifies
+that they are byte-for-byte identical, and checks the rendered files without
+writing.
+
+Edit `docs/source/cuml-accel/benchmarks.rst.in` for narrative or structural
+changes. Do not edit the generated `benchmarks.rst` or SVG files directly.