Features to improve obstacle avoidance functionality

## 🚀 Feature

- Implicitly handle planar or 3D signed distance fields (SDF).
- Support to initialize from occupancy grid (and calculate SDF internally). Will of course be slow for large and/or 3D grids (and we can put a comment for it).
- Flexibility for passing user defined custom obstacle cost functions (possibly overlaps with learnable cost function api).
- Provide option to pass PlanarSDF if its data is constant, to avoid creating a new one for each factor.
- Ensure that map edges are correctly considered: add a flag for providing user options to decide if out of map bounds is free or occupied. (see context below)
- Evaluate the interplay of different options with outer loop optimization (e.g., when learned initial variables go out of bounds during the outer loop optimization).
## Motivation

## Pitch

## Alternatives

## Additional context

https://github.com/facebookresearch/theseus/blob/main/theseus/embodied/collision/signed_distance_field.py#L59


Start with theseus/embodied/collision/signed_distance_field.py around line 59 and trace the current PlanarSDF usage and obstacle-cost flow. Review the existing collision and optimization entry points before deciding how the listed SDF, occupancy-grid, custom-cost, map-boundary, and outer-loop options should interact. Done means the requested behaviors are implemented and their interactions are evaluated.

Written by the indexing model from the issue text.