feat!: Implement coupling of two or more models

[b8raoult]

The new command anemoi-inference couple config.yaml will run two or more models that requires forcing variables from each others. The models can run in different threads, different processes or different MPI ranks. They can use the same GPU (if the fit) or different GPUs.

Below is a sample configuration file, that runs an atmosphere model and an ocean model:

# How to run the model:
transport: processes

# The list of tasks to run
tasks:
  atmos:
    runner:
      config: atmos.yaml
      overrides:
        device: cuda:0
  ocean:
    runner:
      config: ocean.yaml
      overrides:
        device: cuda:1

# List of variables to exchange
couplings:
- atmos -> ocean:
  - lsm
  - 10u
  - 10v
  - 2t
  - 2d
  - ssrd
  - strd
  - tp
  - msl

- ocean -> atmos:
  - avg_tos
  - avg_siconc

[mchantry]

Would be good to develop this multi-GPU with the expected multi-GPU inference for a single model in mind. Perhaps @cathalobrien and/or @japols can have a look at this?

[JPXKQX]

I see the PR has breaking changes. I think it would be helpful for the users if you include a list of the breaking changes in the PR description