Rebuild NILMbench as a reproducible real-data benchmark

[nipunbatra]

Summary

• replace order-dependent benchmark notebooks with an installable nilmbench CLI
• encode typed historical and corrected T1/T2/T3 task profiles in TOML
• use NILM Metadata as the source of truth for appliance taxonomy and synonyms, while recording exact resolved appliance, meter/MeterGroup, AC power type, coverage, and shared-circuit contamination
• separate legacy fixed-10-W F1 from the paper's appliance-specific thresholds
• add deterministic seeds, resumable and identity-safe Optuna studies, structured JSON/CSV results, and source/runtime/container provenance
• add separate CPU-smoke and pinned PyTorch 2.6 / CUDA 12.4 benchmark images
• register PatchTST and pin its exact model commit from Add PatchTST sequence-to-point disaggregator nilmtk-contrib#100
• document discrepancies recovered from the executed Ramanujan notebooks

Protocol findings

The recovered notebooks and paper disagree in several places: REDD T1 asks for data before the canonical file begins; REFIT T1 requests an unavailable microwave; REDD T2 MAE uses building 4 while the F1 notebook and paper use building 6; and notebook F1 uses NILMTK's fixed 10 W threshold although the paper specifies appliance thresholds. Historical profiles preserve those inputs for forensic reproduction. Corrected profiles use the paper splits, strict coverage, per-appliance alignment, and explicit thresholds.

NILM Metadata remains responsible for semantic interoperability. NILMbench adds only experiment-level configuration and provenance; it does not maintain a competing appliance alias table.

Verification

• pytest: 9 passed
• ruff check src tests: passed
• uv lock --check: passed
• docker compose config --quiet: passed
• wheel and sdist: built and inspected
• all eight corrected T1/T2/T3 tasks: passed read-only preflight on checksummed REDD, REFIT, and UK-DALE HDF5 data on Ramanujan
• real REDD PatchTST CUDA smoke on A100-SXM4-80GB: 1 epoch, 1,024 train/test samples, 20.98 s, MAE 64.27 W, F1 0.6675 at the 50 W fridge threshold

Dependency

This PR is intentionally stacked on nilmtk/nilmtk-contrib#100 and pins commit f152f9f1783161bced09da7239d7ad57f4196a9b in the Python lock and container workflow.