[Epic] 4.B GP Pathwise Sampling and Multi-Output Models

## Theme
Add the function-sampling and multi-output GP machinery that scales or broadens the GP modeling surface.

## Parent Wave
- Wave epic: `#33`
- Wave label: `wave:4`
- Milestone: `v0.4-structured-gp-and-uncertainty-aware-nn`

## Motivation
Pathwise (Matheron) sampling provides cheap posterior function draws that moment-based prediction cannot, and the LMC / ICM / OILMM structures extend the GP surface to vector-valued targets — both are prerequisites for the Wave-5 inter-domain and integration work.

## Issues
- [x] #39 — gp(pathwise): PathwiseSampler / DecoupledPathwiseSampler (closed)
- [x] #40 — gp(multi-output): LMC / ICM / OILMM (closed)

## Execution Notes
Both implementation children have landed. Remaining before closing this epic: verify the Definition of Done (pathwise sampling in dense and sparse settings; design-doc multi-output structures represented) and either close the epic or attach follow-ups here.

## Parallelism
- Can run in parallel with: #34 and #36 once prerequisites exist
- Blocked by (inside this wave): none
- Must complete before: Wave 5 inter-domain and integration work that depends on these surfaces

## Definition of Done
- Posterior pathwise sampling works for dense or sparse settings.
- The main multi-output GP structures from the design docs are represented.

## Relationships
- Parent wave: #33.
- Blocked by #22 and #28.