FEAT: validate serialized checksums

[redeboer]

• Closes Validate serialized model checksums with structured results #216
• Closes Make serialized amplitude models easy to load and validate #218

---

✨ New features

• Add validate_checksums(), returning an immutable ChecksumResult for every checksum with its target, point name and values, numeric reference and result, absolute difference, pass state, and optional diagnostic.
• Accept a symbolic Workspace, compiled on demand per checksum target, or reuse an existing CompiledWorkspace.
• Parse real and complex serialized reference values, automatically apply distribution coordinate maps, and infer the symbolic variable for one-dimensional named functions.
• Apply configurable atol and rtol using |value - reference| <= atol + rtol * |reference|.
• Report missing targets, compilation or evaluation exceptions, and non-finite values as failed records instead of assertions or table-specific output.

ChecksumResult and validate_checksums are re-exported from ampform_dpd.io.serialization with no Pandas or plotting dependency.

🐛 Bug fixes

• Compile checksum targets independently so one unsupported or deeply nested target produces its own diagnostic without aborting validation of the remaining targets.
• Map single-coordinate named-function points to the function's actual symbolic variable instead of assuming serialized invariant naming.

🖱️ Developer experience

• Cover real and complex values, NumPy/JAX parity, tolerance boundaries, named-function coordinates, missing targets, compilation failures, invalid evaluations, and non-finite output.

Squash commit messages

* DX: test checksum validation failure paths
* FIX: isolate checksum validation failures