Fix O(N^2) serialization in analysis step

[rhyschappell]

In Analyze.run_parallel, tasks are submitted as ex.submit(self.compute_metrics, i). Submitting a bound method pickles self. That is, the entire Analyze object, including self.data.predict_set, which holds one path per design, gets pickled for each task. This means O(n_designs^2) serialization time, which is expensive in the low 10s of thousands of designs range, and blows up in the high 10s of thousands/low 100s of thousands range.

This PR pickles self once per worker instead of once per task.

I tested this change on a 50k design run and see a per-task pickle size drop from ~9MB to 16 bytes. On a 100k design run this takes the estimated analysis time from ~18hrs to ~1.5hrs.