Relocalization tuning system, wrap autoresearch toolkit for relocalization for new embodiments

Relocalization algo has standard components but heavily depends on embodiment, sensor and sensor mounting spot.

We need a ready to go autoresearch package to run on collected datasets for each sensor style. Keep in mind relocalization is stohastic and chaotic, autoresearch eval needs to be deterministic in order for improvements to be noticable and not random (we had to peg a bunch of random seeds and run a process (not a thread) per core)

https://github.com/dimensionalOS/dimos/pull/2137

Refactor relocalization into an easy to use chainable toolset (ala streams) with tunable hyperparams per processing step

Potentially run an actual classical search on hyperparams before handing off to autotune

(mention in instructions that people shouldn't evaluate relocalization based on the same dataset the relocalization target map was built fromm)

Synced from DIM-944 by summer

Start by reviewing PR 2137 and locating the existing relocalization implementation. Define the scope of the chainable toolkit, per-step hyperparameters, deterministic evaluation process, and dataset guidance before implementation. Done should include a usable autoresearch package for the supported sensor styles with repeatable evaluations.