Allow setting of diffusion in model constructors

[gustavdelius]

Currently newSingleSpecies(), newTraitParams() and newCommunityParams() do not have a diffusion parameter.

[gustavdelius]

I think the right thing is to add D_ext parameter to these functions.

[gustavdelius]

Having traced how diffusion flows through the constructors, here is a concrete plan.

How diffusion already works

External diffusion lives in the ext_diffusion slot, computed from the D_ext species-parameter column as a power law D_ext * w^(n+1) (setExtDiffusion.R). newMultispeciesParams() → setParams() → setExtDiffusion() already reads the D_ext column and defaults it to 0. So if a constructor puts a D_ext column into its species_params data frame, the ext_diffusion slot gets populated automatically — no extra wiring needed.

All three constructors currently build their initial steady state assuming zero diffusion: they pass an all-zero diffusion matrix to get_steady_state_n(), and the docs say "The diffusion rate is set to 0." Note that get_steady_state_n() already accepts and handles a diffusion matrix — it is just being fed zeros.

Proposal

1. Add a D_ext = 0 argument to newSingleSpeciesParams(), newTraitParams() and newCommunityParams().
2. Add a D_ext = D_ext column to the species_params data frame each constructor builds. This alone makes ext_diffusion populate correctly via the existing setParams() path.
3. Feed the resulting params@ext_diffusion (instead of a zero matrix) into the get_steady_state_n() calls in newSingleSpeciesParams() and newTraitParams(), so the constructed initial condition stays at steady state when D_ext > 0. (newCommunityParams() builds its initial spectrum analytically as a pure power law and uses constant reproduction, so it would just set the slot.)
4. Update the docstrings (drop/replace "The diffusion rate is set to 0") and NEWS.md.

This is fully backward compatible: with D_ext = 0 the diffusion matrix is all zeros and behaviour is byte-identical to today.

Two notes:

• In newTraitParams() the post-solve normalisation (mult) uses a power-law assumption, so with large D_ext the initial state would be only approximately steady — the same caveat that already applies to that model.
• D_ext would be a scalar applied to all species (recycled across rows in the trait model), matching how h, beta, etc. are handled.