refactor: derive agent counts from the PDL, one entity one agent (#37)

[Hackatro]

#37 — PDL-literal agent counts (n=1) and consumer split

One PDL entity is now one agent. The n_* count columns are gone from scenario.py and both scenario CSVs, count_attr is gone from topology.py and model.py, and the pooled eu_farmers list is split into poultry_farms, pig_farms and dairy_farms: three roster entries named after their entity ids, exactly like producers and wholesalers. Wholesaler economics are untouched.

Scenario 0, final period

	Before	After
Produced	2300 t	300 t
Delivered to processors	800 t	300 t
soy_price	523.41	673.20
feed_price	1263.00	2618.81
Transport utilisation	0.50	0.14
Producer unit prices (BRA/ARG/US)	396 / 462 / 528	unchanged

Findings

• Bugfix: _step_consumer divided the traders' feed by the agents in its own list, which after the split would have delivered three times the feed produced.
• Prices rise because fixed costs are per-agent and throughput fell tenfold.
• No capacity binds any more. Reasoning, not measurement: Argentina and the US alone offered 1300 t into a Rotterdam lane capped at 500, so a Brazilian drought left delivered volume there unchanged and only the Hamburg branch responded. At n=1 there is no surplus, so a supply shock now passes through the chain proportionally.

Follow-ups being filed

1. Chain constants — capacities, SEA_CAPACITY, base_yield and the fixed_costs_* are calibrated against agent counts that no longer exist.
2. Consumer modelling: consumers ignore PDL impacts aimed at them (livestock_pressure declares supply −9% / price +20% on poultry_farms), and the feed share assumes all consumers share one upstream.