fix: keep the last completed state when an unsaved experiment cycle fails

[ToukoUrsin]

Description

With save_at_cycles, if the solver fails on the first step of a cycle that is not saved, Simulation.solve returns a state from before the last completed cycle, and on_cycle_end can fire for the failed cycle with the previous cycle's summary.

Minimal repro (four 1 s rests, solver fails on the 3rd step):

import pybamm

class FailOnThirdStep(pybamm.IDAKLUSolver):
    calls = 0
    def step(self, *args, **kwargs):
        type(self).calls += 1
        if type(self).calls == 3:
            raise pybamm.SolverError("controlled failure")
        return super().step(*args, **kwargs)

sim = pybamm.Simulation(
    pybamm.lithium_ion.SPM(),
    experiment=pybamm.Experiment(["Rest for 1 second"] * 4),
    solver=FailOnThirdStep(),
)
sol = sim.solve(save_at_cycles=[1], calc_esoh=False)
print(sol.t[-1], sol.all_summary_variables[-1].last_state.t[-1])
# main: 1.0 2.0 (two cycles completed, but the returned state is from cycle 1)
# this PR: 2.0 2.0

Fix. When no step of the failing cycle completed, keep current_solution.last_state (if it is newer than what was retained), reset its solver statistics so discarded work is not counted twice, and stop before the cycle-end callback.

Tests. New tests/unit/test_experiments/test_failure_retention.py (14 cases: list/all/integer save_at_cycles, no failure, first/later failure, resume from a failed sparse solve, partial-cycle failure, solver statistics, two meshes). Six fail on current main, all pass with the fix; tests/unit/test_experiments and tests/unit/test_simulation.py pass.

Related: #4793 (this does not fix that convergence failure; it fixes what is returned after one).

Type of change

Bug fix, with a CHANGELOG.md entry.

🤖 Generated with Claude Code (Claude Opus 5.5).

[rtimms]

Thanks @ToukoUrsin, the diagnosis is right, and the new tests fail on main and pass here. A few small things before merge:

1. Please drop last_state.solver_statistics = pybamm.SolverStatistics() and its comment. Solution._build_last_state already sets last_state's statistics to zero, so the discarded work is not counted twice either way. All 14 tests pass without the line, and it avoids mutating the cached last_state, which SummaryVariables also holds.
2. Please drop test_unsaved_failure_retention_is_mesh_independent. Mesh size isn't involved in this code path, and test_failure_keeps_latest_completed_state already covers the same assertions.
3. CHANGELOG.md conflicts with main; please merge main and keep the entry under # [Unreleased]. Could you also say what changed for the user, e.g. "Simulation.solve with save_at_cycles now returns the last completed cycle's state when the solver fails on the first step of an unsaved cycle, and no longer calls on_cycle_end for the failed cycle."

#5899 changes the same solve and get_data_dict paths, so it'll need a rebase after this and #5724 land.