Add long-form SIR convergence tests covering diverse topologies and parameter regimes

[Copilot]

Existing SIR tests only ran 10–100 steps. Issue #1383 requires coverage of longer-form convergence behaviour across realistic use cases.

New test file: tests/test_sir_long_convergence.py

30 tests across 9 classes, running 200–500 steps against both SIRDynamics and simulate_sir_multiplex_discrete:

• Conservation — S + I + R = N holds throughout 200-step runs on ER, BA, grid, and path graphs
• Monotonic recovery — R[t] is non-decreasing for 300 steps (absorbing-state invariant)
• Extinction — infected count reaches zero by step 500 (SIR always extinguishes)
• Prevalence bounds — prevalence ∈ [0, 1] throughout all long runs
• Phase transition — super-threshold (β ≫ γ, dense graph) grows outbreak; sub-threshold fizzles
• Final-size relation — higher β produces a strictly larger final R
• Multi-replicate consistency — independent seeds all converge to extinction
• Multiplex long-run — simulate_sir_multiplex_discrete conserves population over 200 steps
• Result length — len(results) == steps + 1 for various step counts

# Example: conservation over 200 steps on a BA graph
G = nx.barabasi_albert_graph(80, 3, seed=0)
sir = SIRDynamics(G, beta=0.3, gamma=0.05, initial_infected=0.05)
sir.set_seed(7)
results = sir.run(steps=200)
counts = results.get_measure("state_counts")
N = G.number_of_nodes()
for t in range(201):
    assert counts["S"][t] + counts["I"][t] + counts["R"][t] == N