Normalize scalar types for prior, likelihood, and posterior in Link

[marie-cloet2000]

Fixes #68

Summary

This PR normalizes the types of prior and likelihood in the Link class to ensure they are always scalar values. This prevents downstream failures in diagnostics (specifically to_inference_data) caused by mixing Python scalars
and 1‑element NumPy arrays.

Motivation

While Link.prior, Link.likelihood, and Link.posterior are documented as floats, in practice prior (and therefore posterior) can be passed in as a 1‑element NumPy array depending on the prior implementation. This leads to
inconsistent types:

• prior: np.ndarray with shape (1,)
• likelihood: Python float
• posterior: np.ndarray with shape (1,)

When diagnostics attempt to stack these values into a NumPy array, this results in a ValueError due to inhomogeneous shapes.

Changes

• Explicitly normalize prior and likelihood to scalar values in Link.__init__
• Raise a clear error if non‑scalar values are provided
• Use NumPy‑recommended scalar extraction (.item()) to avoid deprecation warnings

Impact

• Backward‑compatible for existing code
• Preserves current semantics of log‑densities as scalars
• Makes to_inference_data robust to different prior implementations

Notes

I used Microsoft Copilot for suggestions on how to improve the code and to write this PR message.