Add an optional coarse-root pool to FF16 and TF24

[dfalster]

Both models carried one root pool, fine roots proportional to leaf
area, so below-ground structure did not scale with tree size. Coarse
roots are now the below-ground continuation of the sapwood cylinder,
m_cr = a_cr1 * m_s, which makes the structural root:shoot ratio
size-invariant and keeps the growth allocation closed-form.

a_cr1 defaults to 0, so no model's output and no scientific version
moves. expand_state() gains a mass_coarse_root column.

Closes #349

[dfalster]

Why this shape

Three functional forms were on the table for m_cr. Scaling on sapwood mass won for two reasons:

1. It is the only one that scales with size. Fine roots are a_r1 * A_l; a coarse pool of the same form would just be a relabelled a_r1, and the gap [TF24 allometry] Missing coarse root biomass in ff16 model #349 names — no below-ground structure growing with the tree — would still be there.
2. It keeps the growth allocation closed-form. d m_cr / d A_l = a_cr1 * d m_s / d A_l, so darea_leaf_dmass_live gains one addend that reuses the dmass_sapwood_darea_leaf term bark already shares. Scaling on total stem was the alternative and was rejected: heartwood is an ODE state rather than a function of A_l, so mass_live stops being f(A_l) and the allocation derivative needs a full rederivation.

Why a_cr1 = 0 and no version bump

The brief was to bump the FF16 scientific version. That changed once the blast radius was clear: FF16's baselines are not snapshots but a cross-check against the original 2012 R implementation in tests/testthat/FF16_reference/, which asserts parameter-name parity in both directions. A non-zero default would put C++ and the published FF16 equations out of agreement on growth, for a parameter value nobody has calibrated. Shipping the pool off keeps that agreement and leaves the choice of a default to whoever has the data.

Every new term is therefore appended, never interleaved, so the defaults are bit-identical rather than merely close: x + 0.0 == x exactly, and the association order of the existing terms is untouched. (mass_total was the one place I first asserted exactness and shouldn't have — its term order already differs from mass_live's, so the two disagree in the last ULP on develop too. That assertion is now a tolerance.)

FF16@v1, TF24@v9 and TF24f@v9.1 all stand.

The drift guard fired, deliberately

test-model-version.R snapshots each model's default pars/control and failed on the three added names. I re-blessed it after checking the diff is exactly three added entries per model —

	FF16	TF24	TF24f
a_cr1	0	0	0
k_cr	0.2 (= k_s)	0.2	0.2
r_cr	6.5987 (= r_s)	6.5987	6.5987

— with no existing default changed and no version string moved. r_cr/k_cr default to the sapwood rates because coarse roots are woody; as with r_b, that is a default and not a link, so setting r_s afterwards does not move them.

Coarse roots are kept out of TF24's water uptake

TF24 distributes fine-root mass over soil depth to drive uptake and reads a_r1 directly to do it. Absorbing surface is a fine-root property, so coarse roots are excluded by construction, and test-coarse-roots.R pins it: at fixed size, root_mass, transpiration, E_up_, opt_psi_stem, opt_root_psi and assimilation are all bit-identical with the pool on.

Testing

New tests/testthat/test-coarse-roots.R, 22 assertions, split between "it really is inert at the defaults" and "it is correct when switched on" — a default-off feature only ever run at its default is not tested at all.

The load-bearing one reconstructs dH/dt outright from the published chain rather than checking a ratio, because a ratio survives an error common to both denominators:

dH/dt = dH/dA_l * P * (1 - r(H)) / (dM_live/dA_l)

with r_cr = k_cr = 0 so net production is identical on both sides and the whole change is allocation. Agreement to 1e-12 against R formulas written from inst/docs/FF16/FF16-eqns.csv, not read back from the strategy.

Also extended: test-expand-state.R re-runs its oracle comparison with a_cr1 = 0.25 (at the default every coarse-root term is zero, so the existing checks would pass even if C++ ignored a_cr1 entirely), and both AD kernel tests now carry a non-zero a_cr1 so the new terms are actually differentiated.

Full suite: 3221 pass, 0 fail, plus the scenario gateway run separately with PLANT_RUN_SCENARIOS=1 — green, so TF24's end-to-end numbers have not moved.

Note for whoever builds this

develop pins odelia (== 0.3.1) but 0.4.0 is what's installed locally, so compile_dll() refuses before it compiles anything (the #643 blocker). I built against a git archive of the v0.3.1 tag into a scratch library rather than touching the pin.

Not done here

The vertical distribution of root mass over depth — @aornugent's comment on #349, mapping the Rhizomorph cases with an eta-analogue over (depth_max, 0] — is a separate piece of work. This adds the pool; it does not place it in the soil.