tests/ordered: env-driven 2-iteration dry-run cost report + peak-liveset

Add a hidden [.][dryrun-2iter-report] cost-analysis harness over the
water-20 CSV-CCk residual: selects the system (SEQUANT_UT_DRYRUN_SYSTEM =
water20 | c60; adds the C60 pVDZ-F12 size struct) and the batching mode
(SEQUANT_UT_DRYRUN_BATCH = none | aux | aux_occ), runs BOTH forest and the
ordered executor for TWO CC iterations through one persistent cache, and
reports per-iteration builds / FLOPs / peak_bytes (cold iter-1 vs warm
iter-2) plus the {Κ}-block count and the needed_build-skipped BuildStep
count. A SEQUANT_UT_PEAK_LIVESET-gated pass installs on_peak_liveset and
prints the co-resident set at the high-water moment (per-entry GB,
carries-Κ flag, node signature) with all-vs-aux-free totals -- the tool
that showed the aux-only peak is set by aux-free 4-occ 2-PNO integrals.
Report, not a strict gate (sizes are modeled); the ordered aux_occ build
asserts (non-innermost forced split, unimplemented) and is caught
gracefully, forest rows still reported.

Author: evaleev