Fix StridedView sentinels and quadrature regressions from e941630, add ctest target

[gouarin]

Summary

Compile-verified fixes for StridedView and for three regressions introduced in e941630, plus a first ctest target.

StridedView (include/LNIT/misc/StridedView*.hpp)

• isDone() is now const. operator==(const Iterator&, Sentinel) calls it, so any comparison with the sentinel failed to compile once the view was instantiated, and std::sentinel_for / std::ranges::input_range were not satisfied. The header is not instantiated by the library build, which is why this went unnoticed.
• The conditional_t selecting reference was inverted: the non-const iterator yielded const&.
• Random-access advance is bounded by the distance to end() instead of stepping past it when the size is not a multiple of the stride.

Gauss-Laguerre / Gauss-Hermite

• inner_product was given fx.end() instead of fx.begin() (3 sites), reading past the node array. This produced NaN for every moment of order >= 2 in demo_GaussLegendre and a fully-NaN Hessian in demo_LevermooreLikePDF.

AdaptiveQuadratureBase::integrate

• const auto& [a, b] = m_intervals[maxErrIdx] aliased the entry overwritten on the next line, so b became midPoint and the second half of a split interval was integrated over an empty range. The scheme then reported convergence with a wrong value (0.163 instead of 0.5 for the first moment of N(0.5, 1)). Replaced by a copy.

Tests

• New LNIT_BUILD_TESTS option (OFF by default) and tests/test_StridedView.cpp, framework-free, built with the library's -Werror flag set. Covers random-access, bidirectional and forward-only ranges, sizes not multiple of the stride, empty views, mutation through the non-const iterator, sentinel comparison in both operand orders, and static_asserts on the C++20 range/iterator concepts.

Verification

Toolchain: clang (zig 0.16) + libc++, aarch64, fmt 11.1.4.

• Library + 6 demos + test build with zero warnings under the project's -Werror flags in C++20.
• ctest: StridedView passes in C++20 and C++23. The test also passes standalone under ASan/UBSan.
• demo_GaussLegendre: all 101 moments, all 6 tolerances, relative error <= 7e-11, no NaN (before: NaN from order 2, then wrong values after fixing only the fx.end() bug).
• No NaN in any of the 25 demo log files; the Levermore Hessian is now a consistent Hankel-like moment matrix.
• Not verified: full library build in C++23 mode, because ClenshawCurtisAdaptiveQuadrature_impl.hpp uses std::views::stride, which the libc++ available here does not implement yet. Unrelated to this change.

StridedView cleanup (third commit)

• m_bound was computed from std::ranges::distance and never read. Removed. The distance call also consumed pure input ranges before iteration.
• Single-pass ranges have move-only iterators, so Iterator's constructor now takes its arguments by value and moves them, postfix ++ returns a copy only when the inner iterator is copyable (void otherwise), and stride() forwards its argument to std::views::all.
• New test case built on std::views::istream. Clenshaw-Curtis demo logs are bit-identical before and after this change.

Clenshaw-Curtis tables (fourth commit)

• All five CC weight tables (9/17/33 nodes, hybrid 13, GLCC copy of 33) lacked the 1/2 factor on the last Chebyshev mode: exact only up to degree n-1 instead of n+1, and ∫T_n off by exactly a factor 2. Nodes and companion rules were correct.
• New tools/generate_clenshaw_curtis_tables.py (mpmath, 50 digits) regenerates the tables and rewrites the headers in place; the tables now come from it.
• New tests/test_ClenshawCurtisTables.cpp: sum of weights, symmetry, positivity, exactness for x^k up to the declared degree, ∫T_n, and failure at the next even degree. Tables are exposed to tests via a LNIT_TESTING compile definition.
• Shared test infrastructure: tests/Check.hpp, tests/Tolerances.hpp (named tiers), CMake helper with a 30 s timeout.
• estimateIntegralImpl lambdas declared -> LongScalar returned a Scalar: fails -Wdouble-promotion for <double, long double>, which nothing instantiated before. Explicit conversions added (6 headers).
• Effect on demo_ClensawCurtis: max relative error over the 101 moments drops from 1.5e-8 to 7e-10 at tolerances 1e-4..1e-10. At tolerance 1e-14 it goes from 3e-12 to 5e-10 on moments of order 63 to 99 (values 1e44 to 1e79) where the error estimator is optimistic by 7 orders of magnitude. That is an estimator calibration issue, unrelated to the tables, left for a later change.

AdaptiveQuadratureBase state and bounds (fifth commit)

• m_hasConverged was never initialized (indeterminate read on a fresh object). Default member initializers for m_hasConverged and m_it; a private resetState() also runs when an infinite-domain integration gives up on a non-finite tail.
• integrate(f, xmin, xmax) bounds conventions: xmin > xmax returns the opposite of the integral over [xmax, xmin] (before: 0 and "converged" for a short reversed interval, about 4e9 intervals for a long one); equal bounds return 0; a NaN bound returns NaN, not converged (before: plausible value, converged); an infinite bound delegates to the infinite-domain methods.
• New tests/test_AdaptiveQuadratureBase.cpp covering fresh-object state and every bound case for the four adaptive quadratures.

Verification of the last two commits

• ctest: 3/3 pass in C++20; all three tests also pass standalone under ASan/UBSan.
• Library and demos build with zero warnings under the -Werror flag set.
• C++23: only test_StridedView could be built here, the library needs std::views::stride which the available libc++ lacks.