Tests: split test_numeric_lu per element type (CI critical path)

TASKS.md

@@ -2577,7 +2577,8 @@ uniform-batch LU tests pass; `random_general` residuals `≤ tol(T)`.
   - #84 (found-by-agent): views `'L'`/`'U'` with `"G"`.
   - Regime-C performance (vendor BLAS-3, parallel pivot search) with #75 in T25. The pivot search is serial on
     work item 1 (`O(f)` per column, cheaper than the LDLᵀ fallback scan).
-  - `perm_row` reports batch member 1 for a uniform batch (each member has its own local row order).
+  - #92 (found-by-agent): `perm_row` reports batch member 1 for a uniform batch (each member has its own local
+    row order).
   - CUDA path exercised only by CI.
   - Device arithmetic fuses multiply-adds (GPUCompiler `-nvptx-fma-level=1`), so device-vs-reference equality is
     bitwise on the CPU backend only. On `weak_diagonal_general` with `pivot_threshold = 0.01` (growth ≈ 3e2) CUDA

test/runtests.jl

@@ -45,8 +45,8 @@ end
 # worker compiles the solver for every element type it tests, and that compilation, not the tests, sets the wall
 # time. The testsets that do not loop over the element types run in the Float64 part (`RUN_SHARED`).
 const SPLIT_FILES = ["test_api", "test_dense", "test_fgmres", "test_numeric_cholesky_a", "test_numeric_cholesky_b",
-                     "test_numeric_cholesky_c", "test_numeric_ldlt", "test_ported", "test_refinement", "test_solve",
-                     "test_ubatch"]
+                     "test_numeric_cholesky_c", "test_numeric_ldlt", "test_numeric_lu", "test_ported", "test_refinement",
+                     "test_solve", "test_ubatch"]
 const SPLIT_ELTYPES = (Float32, Float64, ComplexF32, ComplexF64)
 
 # the seed goes with the test, not into `init_code`: ParallelTestRunner seeds with 1 after `init_code`