Implement fused kernels for vegetation carbon compute_auxiliary!

[bgroenks96]

This PR adds kernel fusion for vegetation carbon processes. Description below will be updated after those changes are added.

The following is generated by an LLM 🤖

Fuse the vegetation carbon cycle auxiliary kernel

This PR fuses the vegetation carbon cycle compute_auxiliary! chain into a
single kernel launch, and documents (via a reverted experiment) why the same
fusion does not work for surface hydrology.

What changed

Vegetation carbon cycle fusion (kept). The five-stage chain

carbon dynamics → phenology → photosynthesis → stomatal conductance → autotrophic respiration

previously ran as five separate compute_auxiliary! launches. Each stage is a
pure producer/consumer of the next within a cell (balanced_leaf_area_index →
leaf_area_index/phenology_factor → net_assimilation/
gross_primary_production → canopy_water_conductance → net_primary_production),
so the whole chain now resolves inside one XY launch
(compute_auxiliary_kernel!). Root distribution and plant-available water still
run beforehand as separate launches: they depend on the soil state and produce
fields (soil_moisture_limiting_factor) that the fused kernel reads.

Surface hydrology fusion (tried, reverted). The same fusion was implemented
for SurfaceHydrology (interception → {evapotranspiration, surface runoff}) in
d39a54c and reverted in afdb0a5. Unlike the vegetation chain,
evapotranspiration and runoff are independent siblings, not a dependency
chain: fusing them forces the register union of two unrelated stages, which
regressed GPU performance by ~20% with no end-to-end benefit. The per-process
fan-out is restored byte-for-byte. This finding is recorded in the plan doc
(rev 7) as the boundary of the fusion rule: fuse dependency chains and single
accumulators, do not bundle independent outputs.

Docs fix. The @docs signature for
compute_auxiliary!(::PALADYNCanopyEvapotranspiration, ...) had gone stale
(it listed a snow parameter that the method does not take), which broke the
Documenter build on this branch. Fixed in 955111b.

Benchmarking

Microbenchmark of the vegetation carbon-cycle auxiliary pass (fused single launch vs the
pre-fusion seven-launch fan-out), interleaved A/B, minimum of 30 repeats, Float64, nz=30,
on the :land configuration at three resolutions:

nlat_half	columns	CPU speedup (min)	A40 GPU speedup (min / median)
8	512	1.00x	1.01x / 1.02x
24	4,608	1.01x	1.06x / 1.03x
48	18,432	1.00x	1.01x / 1.01x

Neutral on CPU and neutral-to-slightly-positive on GPU, as expected for a genuine
dependency chain: the fused kernel resolves the chain in one launch without paying the
register-union penalty that independent-sibling fusion incurs (see the reverted surface
hydrology experiment, ~20% GPU regression).

Testing

• New integration test (test/vegetation/integration_tests.jl): the fused
  kernel's auxiliary output (leaf_area_index, phenology_factor,
  net_assimilation, gross_primary_production, canopy_water_conductance,
  net_primary_production, ...) matches the per-process fan-out exactly,
  across the vegetation carbon-cycle configurations.
• Existing vegetation, surface hydrology, and land model suites pass.

Files

File	Change
src/processes/vegetation/vegetation_carbon_cycle.jl	Fused compute_auxiliary_kernel!, rewritten compute_auxiliary!
test/vegetation/integration_tests.jl	New fused-vs-fan-out equivalence test
test/vegetation/vegetation_model_tests.jl	Include the new test file
docs/dev/2026-08/2026-08-31-PLAN_fused_kernels.md	Rev 7: scope narrowed to vegetation; surface/soil fusion findings

[maximilian-gelbrecht]

Looks pretty straight forward to me, just one question about the output field assignment (see comments)