Improve concurrency design

[chlowell]

There are a number of scenarios in which waza isn't as concurrent as it could be or appears it should be given configuration. For example, waza run --workers 6 ... on an eval suite having 3 tasks and 3 trigger tests gets only 3 concurrent sessions because all tasks complete before the first trigger test runs. My first thought is, how about a worker pool?

[richardpark-msft]

We'll also get some mileage out of simply treating the copilot SDK client as a process-wide singleton. That wasn't super practical before, but now that we've pushed all the unique state to sessions (workdir, model, etc..) we can do it.

I'm not entirely sure where we're bottlenecking, but it'd be great to make it faster for local cases.