v0.4.0

• Refactor sweep across the Python and Rust sources (net -400 lines): dedupe
  the GRAPE entry-point validation preamble, the grape_xy basis dispatch,
  the Hessian generator assembly, and the Rust eigenpropagator; unify
  checkpoint serialisation and waveform validation behind single authorities.
  Behavior verified equivalent against the pre-refactor tree (fixed-seed
  value/gradient/Hessian/fidelity/ensemble/IO comparisons, bit-identical on
  the pure Python path).
• Remove dead public surface: the unused gradient_ascent optimizer, the
  never-wired print_iteration_table diagnostics helper, and the Rust
  grape_member_value_gradients_vectors kernel with its Python wrapper.
  run_grape continues to offer lbfgs and newton.
• Consolidate the ten per-example Bruker shape writers into
  optimalcontrol.io.export_bruker_shape (byte-identical output; the shared
  writer rejects mismatched amplitude/phase lengths).
• Tighten import_jcamp_dx: files must carry ##CHANNELS and
  ##$OPTIMALCONTROL_TIMES (the layout-guessing fallbacks were untested and
  unreachable from export_bruker round trips).
• Restore the shape-mismatch ValueError in final_fidelity that the
  fidelity-dispatch cleanup briefly dropped, with a regression test.
• Fix plotting helpers to resolve the parent figure on any matplotlib
  version instead of requiring 3.10 for get_figure(root=True).
• Make mypy actually check the package (the pinned python_version = 3.10
  aborted inside numpy stubs on Python 3.14) and fix the seven latent typing
  errors it revealed.