V3.6: Demo notebook — twin advection on Gulf Stream uo/vo

## Summary
End-to-end notebook tying V3.1–V3.5 together. Twin experiment on a real CMEMS surface velocity dataset on the Gulf Stream `[-70, -50] × [30, 45]`, plus comparison to an actual drifter passing through the region.

## What to build
`docs/notebooks/validation_v3_lagrangian.ipynb`:

1. Load CMEMS surface `uo` / `vo` (e.g. `cmems_obs-mob_glo_phy-cur_my_multi_P1D`) for ~30 days.
2. Make a "perturbed" version of the velocity (low-pass filter + small bias) as the "model" twin.
3. `SeedParticles` — grid of seeds upstream of Cape Hatteras.
4. `AdvectParticles` on truth + perturbed; visualize the trajectory bundle.
5. `PairDispersion` → curves diverge over time.
6. `ResidenceTime`, `ConnectivityMatrix` over coastal / shelf / offshore regions.
7. `EndpointError`, `DispersionError`, `ResidenceTimeError`, `ConnectivityError` — quantify the twin.
8. Optional: ingest a GDP drifter that crossed the region during the period via V3.5; compute `EndpointError` against both truth and perturbed model.

## Acceptance criteria
- [ ] Notebook executes end-to-end.
- [ ] Inline figures committed (no separate PNGs).
- [ ] `xr_toolz_venv` kernel.
- [ ] mkdocs nav under "Validation tutorials".

Review the V3.1–V3.5 components and the requested entry points—SeedParticles, AdvectParticles, PairDispersion, ResidenceTime, ConnectivityMatrix, and the error metrics—before building docs/notebooks/validation_v3_lagrangian.ipynb. Run it with the xr_toolz_venv kernel against the CMEMS Gulf Stream data. Done means end-to-end execution, inline figures, and a mkdocs nav entry under “Validation tutorials”.

Written by the indexing model from the issue text.