ENH: Native Simpson, deprecate vnl integral classes, retire netlib quadrature

[hjmjohnson]

Reimplements vnl_simpson_integral with a native composite-Simpson loop, deprecates it and vnl_adaptsimpson_integral behind ITK's three-state legacy guard, and retires the now-unused v3p/netlib/mathews/* sources. Advances the VNL→native numerics direction (#6403, #6230); follows the deprecation pattern of #6441 and #6454.

What changed (3 commits)

Commit	What
COMP: Reimplement vnl_simpson_integral	Computes the composite Simpson rule (Mathews Alg. 7.2) directly in C++; drops v3p_netlib_simpru_. Loop bounds/accumulation order match the Fortran exactly → bit-identical results.
COMP: Retire unused netlib/mathews	Removes simpson.*, trapezod.*, adaquad.f, the standalone netlib_integral_test, and the mangle/unmangle/prototype/CMake/README entries. trapru_/adaquad had no caller; simpru_ was the only live one.
ENH: Deprecate both integral classes	__has_include(<itkConfigure.h>)-gated three-state guard (silent default / ITK_LEGACY_REMOVE warning / ITK_FUTURE_LEGACY_REMOVE error), inert for the upstream VXL build. test_integral.cxx defines ITK_LEGACY_TEST.

vnl_adaptsimpson_integral was already pure C++ (it never used the Fortran).

Verification

• ctest -R test_integral passes against the native engine (analytic checks to 1e-6).
• All three guard states verified (default silent / ITK_LEGACY_REMOVE warns / ITK_FUTURE_LEGACY_REMOVE errors); +ITK_LEGACY_TEST opt-out keeps the in-tree test silent.
• pre-commit run --all-files clean.
• Phase-5 gate: native ≤ netlib in both accuracy and time across 5 integrands (bit-identical error; faster on the oscillatory case). Full report posted as a comment below.

[hjmjohnson]

Phase 5 performance proof — native Simpson vs netlib simpru_

Verdict: PASS the gate. The native composite-Simpson loop is an exact
arithmetic transcription of the f2c simpru_ (identical loop bounds, identical
accumulation order, identical final expression), so its error is bit-identical
to netlib in every measured case (native ≤ netlib accuracy holds with equality),
and its median time is ≤ netlib in every case — notably faster on the
oscillatory integrand. Backing the kept vnl_simpson_integral API with the
native engine and retiring the netlib/mathews sources is justified.

Two-axis benchmark (relative error + median ms, lower = better)

case	err_native	err_netlib	ms_native	ms_netlib	Eigen/native ≤
x/(1+x²) [0,1], n=50	1.202e-09	1.202e-09	0.0001	0.0001	both
x/(1+x²) [0,1], n=5000	9.610e-16	9.610e-16	0.0084	0.0084	both
gaussian [-6,6], n=500	0.000e+00	0.000e+00	0.0025	0.0025	both
gaussian [-6,6], n=50000	3.012e-15	3.012e-15	0.2520	0.2519	both
sin(50x)·e⁻ˣ [0,3], n=20000	1.095e-12	1.095e-12	0.2312	0.3816	both

Error = |approx − exact| / |exact| against analytic references
(½ln2; √(2π)·erf; closed form for the oscillatory case).

Environment & reproducer

• OS: Darwin 25.5.0 arm64 — Apple M3 Max — Apple clang 21.0.0
• Single-threaded; median of 15 reps per case; -O2 -std=c++17
• Both engines exercised in one process on identical integrands (fairest
  same-compiler/same-CPU/same-integrand-cost comparison); netlib arm is a
  verbatim copy of v3p/netlib/mathews/simpson.c:simpru_.
• Reproducer: /tmp/simpson_bench/bench.cxx, built with
  c++ -O2 -std=c++17 bench.cxx -o bench, run with
  ITK_GLOBAL_DEFAULT_NUMBER_OF_THREADS=1 ./bench.

[hjmjohnson]

@dzenanz TLDR; These functions are not used in ITK or any of the downstream BRAINSTools/ANTs/Remote Modules/Slicer/SlicerExtentions code bases. It is deprecated immediately, and only here in the very unlikely case of a downstream user who may have needed it. NOTE: The re-implementation was a trivial rewrite in C++.

No ITK-dependent tool that uses these functions. 2 of the 3 functions were provably unreachable or not compiled.