tech-debt: v3.5 docs refresh, ANI2xt fp64 energies, optimizer benchmark

[isayev]

Closes the remaining workstreams from the two post-hardening plans (W7–W9 of the tech-debt roadmap; both plans were otherwise already implemented across PRs #93/#94).

Changes

• Docs refresh for v3.5 (W7) — removed all documentation of the deleted ensemble feature (AUTO3D_USE_ENSEMBLE, --use-ensemble, "8-model ensemble") across 10 doc pages; fixed version floors (Python 3.11, PyTorch 2.8); documented aimnet as a core dependency, the [ani]/[ase] pip extras, the ~/.cache/aimnet model download/cache (AIMNET_CACHE_DIR), and the AIMNet2 registry model names. Sphinx build is clean (0 warnings).
• ANI2xt fp64 accumulation (W8) — accumulate the ANI2xt forward's per-atom and self-atomic energies in float64 so the float64 energy_shifts buffer is no longer silently truncated. Cosmetic: self-atomic shifts cancel in conformer energy differences, so ranking is unaffected; energy is returned as float64 and forces as float32, matching the AIMNet2 adapter contract. The fp32 network output still caps usable precision.
• Optimizer benchmark harness (W9) — scripts/bench_optimizer.py, an opt-in micro-benchmark (outside tests/, never gates CI) that times optimizing(...).run() on a fixed, deterministic drug-like conformer set and reports wall time, throughput, and peak memory. Documented in CONTRIBUTING.
• Small: clarified the @ tautomer-separator docstring in results.py; recorded the two plan docs; bumped a stale "Python 3.10" mention in CONTRIBUTING to 3.11.

Notes for review

• W8 deviates from the written plan, on purpose. The plan's "build atom_energies in float64" crashes at runtime (Index put requires source and destination dtypes match — fp64 destination, fp32 network output). This PR adds the explicit elem_energies.to(torch.float64) cast the plan omitted; verified via a standalone autograd reproduction (fp64 energy / fp32 forces, finite grads). ANI2xt cannot run without the optional torchani extra, so this path must be confirmed by the CI [ani] job — it is not exercised by the local suite.

Verification

• ruff: clean (src/, scripts/)
• Sphinx docs build: clean, 0 warnings
• Test gate (pytest -m "not slow"): 631 passed, 8 skipped (torchani-gated), 45 slow deselected