[Wave 5] Advanced Algorithms and Ecosystem Integrations

## Goal
Complete the advanced algorithm and ecosystem surface: additional GP inference/likelihood methods, inter-domain features, external integrations, and the more ambitious NN layer families.

## Why This Wave Exists
Wave 5 is where the backlog opens into the long-tail research and interoperability surface that makes sense only after the earlier foundations are in place.

## Wave / Milestone
- Wave: `wave:5-advanced`
- Milestone: `v0.5-advanced-algorithms-and-integrations`

## Issues

Theme epics (native sub-issues of this wave), with their open leaf issues:

### 5.A GP advanced inference and likelihoods — #44
- [ ] #48 — `MultiLatentLikelihood` wrapper for chained / heteroscedastic / multi-class GPs

### 5.B GP inter-domain features and integrations — #45
- [ ] #50 — `pyrox.gp.metrics` + GPflow / SPDE-FEM / Falkon integration shims

### 5.C NN advanced layer families — #46
- [ ] #51 — Edward2-style advanced output heads (blocked by jejjohnson/geonnax#12)
- [ ] #52 — Bayesian convolutional / recurrent / attention layer families

## Sequential Dependencies
- All three tracks are mostly parallel after Wave 4.
- Cross-epic example or docs work may depend on earlier advanced surfaces, but the implementation themes should remain separate.

## Definition of Done
- Advanced GP inference and likelihood methods are represented beyond VI/CVI.
- Inter-domain features and ecosystem integrations exist as either code or honest, fully-contained docs/examples.
- The NN track includes the advanced Edward2-style and sequence/vision layer families plus richer tutorial material.

## Relationships
- Blocked by #33.