`BatchedInverseProblem` with asynchronous, device-aware `forward_map` + concatenation

We need a way to run batches of forward simulations asynchronously, and then to concatenate the forward maps deterministically for `EnsembleKalmanInversion`. One way to do this is to develop a `BatchedInverseProblem` that consists of

1. A tuple / array of `InverseProblems`, each either their own `observations` and `simulation`;
2. A utility that concatenates the `forward_map` / `inverting_forward_map` from each individual `InverseProblem` to pass to `EnsembleKalmanInversion`.
3. The ability to extract each `inverting_forward_map` asynchronously: https://docs.julialang.org/en/v1/manual/asynchronous-programming/

We probably also want to make `inverting_forward_map` "device aware", so that we can run simulations on different GPUs on the same node (for example). This won't be hard, since it's just a matter of "switching" to the appropriate device before running any GPU code. We can copy data to the CPU in `FieldTimeSeriesCollector` while the simulations are running, so none of the rest of the code needs to care about this. See CUDA.jl docs or here: https://juliagpu.org/post/2020-07-18-cuda_1.3/index.html.

All of this is relatively simple to implement in that it won't take many lines of code once we know what to write.

Start by reading the existing InverseProblem and EnsembleKalmanInversion entry points, then inspect forward_map, inverting_forward_map, and FieldTimeSeriesCollector. The proposed work is done when BatchedInverseProblem supports deterministic concatenation, asynchronous extraction, and device-aware simulation as described; the issue names no specific files or tests to run.

Written by the indexing model from the issue text.