perf(forest): thread the reference BlockForest path — leaf sweeps and exchange descriptor loops

[kylebeggs]

Carved out of #49 (candidate 1) so it can be tracked as one unit of work.

_forest_sweep_leaves! (src/operators/forest.jl), _forest_adjoint_sweep_leaves!, and the three descriptor loops in _exchange_storage! (src/transfer.jl) are plain serial for loops; on CPU the packed kernels fall back to the same broadcasts. Leaves are disjoint and the exchange phases are ordered to be independent within a phase, so this looks close to embarrassingly parallel. In the Niederer benchmark the reaction term is threaded in example code while diffusion is not, which is a large part of why the two phases scale so differently (#49: diffusion 0.44× on 5.3× fewer cells).

Check first

Whether descriptors within one exchange phase can alias the same destination block. If they can, the threaded sweep needs a partition by destination or atomics; if they cannot, a Threads.@threads over descriptors is enough.

Constraints

• Same executor decision as meta(cpu): CPU apply! path is single-threaded — Julia threads give zero speedup #51 (KA CPU backend vs a threaded executor at the apply! seam); pick once for both single-grid and forest paths.
• Must not disturb the _exchange_storage! / _bc_storage! rule seams: no allocation inside them, no Field-typed rule arguments.
• Acceptance: the examples/niederer_benchmark.jl phase table shows the diffusion phase scaling with -t, and the adjoint sweep keeps ⟨Lx,y⟩ = ⟨x,Lᵀy⟩ bit-for-bit against the serial path.

---

_{🤖 Beep boop — Claude split this off #49; Kyle is not forking himself.}