FEAT: compile serialization workspaces

[redeboer]

Closes #215

✨ New features

• Add compile_workspace(), which compiles selected symbolic workspace functions and distributions through TensorWaves, and the frozen CompiledWorkspace container for their callables, backend, coordinate maps, and selected independent coordinates.
• Select named functions or distributions with targets; by default compile checksum targets, falling back to all workspace targets when no checksums exist.
• Support lazy numpy and jax backends with actionable errors for unsupported backends, missing optional dependencies, and unknown targets.
• Compile each selected distribution's serialized coordinate transformation with model masses substituted.
• Allow callers to choose two independent Mandelstam coordinates and apply parameter_overrides without mutating the symbolic workspace.
• Keep numerical imports and backend state outside Workspace, AmplitudeModel, and symbolic kinematics objects.

CompiledWorkspace and compile_workspace are re-exported from ampform_dpd.io.serialization.

🐛 Bug fixes

• Reduce each distribution to its two independent Mandelstam invariants before compilation, preventing dependent-invariant inputs from being required at runtime.

🖱️ Developer experience

• Cover named functions and distributions on NumPy and JAX, compiled coordinate maps, inferred and explicit invariant coordinates, parameter overrides, and invalid backend, target, and coordinate selections.

📝 Documentation

• Disambiguate public CompiledWorkspace references for Sphinx API links.

Squash commit messages

* DOC: disambiguate CompiledWorkspace in API docs
* DX: test compiled workspace on NumPy and JAX
* FIX: reduce distributions before compilation