Update cuml.accel benchmark results and presentation

docs/benchmarks/cuml-accel/README.md

@@ -0,0 +1,38 @@
+# Synchronizing the cuml.accel benchmark page
+
+The checked-in `benchmark-data.json` is the compact input for the Sphinx page.
+It is produced by `cumlbench-dash`; raw benchmark observations stay outside
+this repository. The file stores the benchmark system and package versions
+alongside case labels, shapes, median timings, CPU timeout limits when
+applicable, and PCA component counts that cannot be derived from labels.
+Speedups, classifications, summaries, input sizes, and display units are
+derived while rendering.
+
+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 accepts publication schema version 1. The artifact contains
+168 cases: the selected 165-case performance grid plus three additional
+medium-wide PCA component-rank measurements. The existing medium-wide PCA case
+supplies the rank-1,024 point. The command validates and copies the artifact
+unchanged, then renders the 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.