refactor: remove demand-side mechanisms superseded by market response

[koen-vg]

Stacked on #48 (rebased onto its final state, including the capacity-lift refit and the legacy-chain restructure). Removes the legacy demand-side mechanisms that the market_response demand component supersedes, migrates the remaining consumers, and rewrites Tutorial 2 around demand response.

Removed

• Consumer-values workflow (extract_consumer_values, the consumer_values.smk rules, validation, comparison plot, docs page): hand-rolled PMP phase 1 for the demand side, with the level-degeneracy problem the sequential calibration was built to fix.
• food_utility_piecewise (+ calibrate_food_utility_blocks): piecewise utility from consumer values with ad-hoc decline/width knobs -- the demand curves carry the same content with exact calibration and literature elasticities.
• food_incentives and the optimal_taxes workflow: disabled everywhere; the equivalent quantities are directly readable off the demand curves.

~1.9k lines of workflow code plus the config blocks, schema entries, solve-time prefixes, and manifest wiring. The deviation_penalty L1 machinery (and the legacy cost + stability steps) stay, as the anchor for coarser curve granularities and structurally divergent configs.

Migrations

• gsa.yaml and doc_figures.yaml complete their move onto the demand component (started in Replace L1 stability anchoring with exactly calibrated PMP market-response curves (supply + demand) #48): the doc-figure baseline scenario and the consumer-values figure are gone, as is the GSA's locally-pinned-solve special case for cluster runs.
• gsa_fixed_diet.yaml keeps components.demand: false + enforce_baseline_diet: true (the schema rule from Replace L1 stability anchoring with exactly calibrated PMP market-response curves (supply + demand) #48 requires exactly one of the two).
• Cost calibration: step 2 keeps the diet pinned to baseline (matching step 1) instead of valuing a free diet with utility blocks, so the tight-band duals measure the cost of holding observed production under the observed diet. The legacy stability calibrator's main solve runs the base config's own demand regime -- for legacy configs that is a free, unvalued diet, matching how they consume the artefact.

Artefact refits

• default and gbd-anchored: the legacy cost and stability artefacts are refit under the new regime (the four-step chain from Replace L1 stability anchoring with exactly calibrated PMP market-response curves (supply + demand) #48, including the capacity-lift market-response fit, is untouched). The calibrated L1 costs rise substantially (default: cropland 1.38 -> 6.54, grassland 0.23 -> 0.34, feed 0.06 -> 0.50 bn USD per unit deviation): without any demand-side valuation, the free-diet main solve reorganizes production much harder, and holding the 5% deviation target is genuinely more expensive. Legacy consumers run the same regime, so calibration and consumption match; expect stiffer L1 behaviour in legacy configs.
• tutorial-02 (new): backs the demand-response tutorial at 200 regions, fit with the four-step chain only -- no legacy cost/stability artefacts ship with it. Config validation now requires the legacy stability yaml only when a solve would actually resolve the "calibrated" sentinel, which replaces the earlier seeding workaround.

Tutorial 2 rewritten

Now "Demand response: letting the diet adjust to policy": the reader calibrates a dedicated 200-region artefact set (tools/calibrate --base config/tutorial/02_demand_response.yaml, committed as tutorial-02), sees the exact-calibration property live (an unpriced solve reproducing the observed food system with nothing pinning it), then runs a carbon-price ladder plus two demand-side variants: the diet frozen at baseline (what flexibility is worth) and half elasticities (how much hinges on that assumption, changeable with no refit). The companion notebook compares consumption shifts, the supply_response/demand_response objective columns, and emissions across the three demand regimes.

Verification

• 758 unit tests pass (including new coverage of the stability-yaml existence gate); DAG dryruns pass for the tutorial, gsa.yaml, gsa_fixed_diet.yaml, and doc_figures.yaml.
• tools/calibrate --check is fully green for all three artefact sets; both legacy stability calibrations converged (default: 4 warm-started iterations, residual 0.016; gbd-anchored: residual 0.013).
• Tutorial reference solve (unpriced, 200 regions, HiGHS, tutorial-02 set): baseline reproduction with nothing pinning it: consumption 0.003%, crops 0.107%, pasture 0.012%, animal feed 0.249% off observed 2020. Demand censoring minimal: 10 of 8403 groups, total pin slack 0.14 Mt.