Refactor configs (hydra), training and evaluation

[Phil26AT]

Large refactor PR that simplifies the existing codebase and adds new functionality. This PR is backward compatible with the initial release.

Main changes:

• Simplified training, better logging
• Much faster training, working compilation: Now training without caching is equally fast as training with caching before (and even faster with caching now)
• Switch from omegaconf to hydra. This enables us to avoid massive code duplication (in configs), which is error-prone. Old configs still work.
• Improve profiling (trace and memory), log important statistics to tensorboard.
• Reorganize checkpoints: Store test results with checkpoint instead of outputs/results.
• Small bugfixes.
• ETH3D SfM eval, based on hloc
• ETH3D 2-view dataset (relative pose + depth)
• Air-Ground Eval
• More to come