Vectorize _apply_impacts loop over realizations

[benmsanderson]

Summary

MeteorInterface._apply_impacts was looping over realizations in Python:

for i in range(n_realizations):
    result = dd_model.calculate(temp_celsius[i])
    hdd_results.append(result.data["annual_hdd"].values)
    cdd_results.append(result.data["annual_cdd"].values)

The underlying DegreeDaysCalculator.calculate is already fully vectorized across the month dimension via xarray, so passing the entire (n_realizations, n_month) DataArray in one call produces identical results without the per-realization Python overhead.

Correctness

• New test_degree_days_calculate_batches_realizations pins the contract: batched output and looped-per-realization output agree to atol=1e-9.
• All existing 46 tests in tests/unit/impacts/ + tests/unit/test_meteor_interface.py still pass.

Wins

On the gen_impacts profiling workload at N=100 (via the sibling profiling harness):

• DegreeDaysCalculator.calculate cumulative was 20.5 s of a 63 s workload — the serial loop was called 100 times per variable × 2 variables = 200 times.
• After this change it's a rounding error in the profile.

Savings scale linearly in n_realizations.

Independence

Independent of the other perf/fix PRs in flight (#91–#95). Touches _apply_impacts (around line 2089) — well away from any other in-flight edits to meteor_interface.py. Can land in any order.

Test plan

• pytest tests/unit/impacts/test_degree_days.py::test_degree_days_calculate_batches_realizations
• pytest tests/unit/impacts/ (full impacts test suite still passes)

🤖 Generated with Claude Code

[maritsandstad]

Fine this can probably just go in