Add composition and composition+persistent-homology benchmark submissions

[nkwork9999]

Summary

Two new algorithm folders under benchmarks/, both scored on the official
Matbench v0.1 fold IDs and metrics:

• matbench_v0.1_composition_gbm — matbench_glass (glass-forming
  ability), composition-only features (156-dim: element fractions + global
  composition statistics) with HistGradientBoostingClassifier. Mean
  official-fold ROC-AUC 0.8658 (± fold std ~0.017).
• matbench_v0.1_composition_ph_gbm — matbench_log_gvrh and
  matbench_log_kvrh (shear/bulk modulus), the same composition features
  plus a 51-dim persistent-homology structure descriptor (H0/H1
  Vietoris-Rips diagrams on the periodic point cloud, via ripser+persim)
  with HistGradientBoostingRegressor. Mean official-fold MAE 0.0813
  (log_gvrh) and 0.0636 (log_kvrh); the PH features improved over a
  composition-only control in all 5 folds on both tasks.

Methodology notes

• All three tasks use the official matbench_v0.1_validation.json fold IDs
  (the local environment could not build matbench==0.6's pinned
  scikit-learn==1.0.1, so results were constructed to the
  MatbenchBenchmark JSON schema directly from the raw datasets + official
  fold JSON — same fold IDs, same metric functions).
• Hyperparameter selection used only an inner split inside each official
  training fold (never the fold's test data); the selected model was refit
  on the complete official training fold before recording test predictions.
• Each folder's src/ is standalone and was verified to reproduce the
  recorded results.json.gz exactly (byte-identical per-fold predictions
  and scores) by running fresh from inside that directory alone.
• Full details, feature definitions, and reproduction commands are in each
  folder's info.json and notebook.md.

Test plan

• results.json.gz in both folders validated against the
  MatbenchBenchmark schema (gzip integrity, official fold IDs/order,
  metric keys, finite predictions).
• src/ reproduction verified standalone (fresh dataset download +
  refit + score) to match the committed results.json.gz exactly for all
  three tasks.
• Maintainer review / merge — flagging that I could not get the
  matbench package itself to build locally (see above), so if the schema
  needs any adjustment for merge I'm happy to iterate.