Update EquivariantTensors to 0.4.2 and improve ET pair memory efficiency

- Bump EquivariantTensors compat to 0.4.2 in main and test Project.toml
- Simplify site_basis_jacobian to use 5-arg _jacobian_X API (requires ET >= 0.4.2)
- Improve ETPairModel site_grads memory efficiency:
  - Avoid O(nnodes * nbasis) intermediate matrix allocation
  - Compute edge gradients directly using loops

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

Author: jameskermode

Project.toml

@@ -65,7 +65,7 @@ ConcreteStructs = "0.2.3"
 DecoratedParticles = "0.1.3"
 DynamicPolynomials = "0.6"
 EmpiricalPotentials = "0.2"
-EquivariantTensors = "0.4"
+EquivariantTensors = "0.4.2"
 ExtXYZ = "0.2.0"
 Folds = "0.2"
 ForwardDiff = "0.10, 1"

src/et_models/et_ace.jl

@@ -72,7 +72,8 @@ end
 function site_basis_jacobian(l::ETACE, X::ET.ETGraph, ps, st)
    (R, ∂R), _ = ET.evaluate_ed(l.rembed, X, ps.rembed, st.rembed)
    (Y, ∂Y), _ = ET.evaluate_ed(l.yembed, X, ps.yembed, st.yembed)
-   # _jacobian_X for SparseACEbasis takes (basis, Rnl, Ylm, dRnl, dYlm) - no ps/st
+   # _jacobian_X for SparseACEbasis takes (basis, Rnl, Ylm, dRnl, dYlm)
+   # Requires EquivariantTensors >= 0.4.2
    (𝔹,), (∂𝔹,) = ET._jacobian_X(l.basis, R, Y, ∂R, ∂Y)
    return 𝔹, ∂𝔹
 end

test/Project.toml

@@ -37,5 +37,5 @@ Zygote = "e88e6eb3-aa80-5325-afca-941959d7151f"
 ACEpotentials = {path = ".."}
 
 [compat]
-EquivariantTensors = "0.4"
+EquivariantTensors = "0.4.2"
 StaticArrays = "1"