feat: release Tinydiffeq 2.4.0

[jlperla]

Summary

• release Tinydiffeq 2.4.0 with exact fixed-step knot output, logical work counters, and an opt-in actual-work adaptive loop
• tighten semi-explicit DAE root acceptance, delegate square implicit AD to nlls-gram, add a secant predictor, and preserve float32 through primal/JVP/VJP lowering
• harden Krylov happy-breakdown AD, Markov RNG dtype handling, fixed-step time arithmetic, documentation, and release packaging
• test Python 3.11-3.13 in CI and validate wheel/sdist metadata

Validation

• CPU: 306 passed, 15 skipped
• GPU: 15 passed
• independent high-risk JAX review: 187 passed; no JAX, AD, numerical, or dtype blockers
• ruff check, ruff format --check, git diff --check, and uv lock --check
• mkdocs build --strict
• clean committed worktree: uv build and uvx twine check dist/*
• isolated wheel smoke test confirmed Tinydiffeq 2.4.0 and a float32 RK4 solve

The experimental July 31 benchmark programs and results remain intentionally untracked and are absent from the release artifacts.