Add fine-tuning guidance and a MACE fine-tuning tutorial

[vue1999]

This PR adds practical guidance for fine-tuning MACE foundation models and a
worked tutorial based on the results from Fine-tuning MLIP foundation models:
strategies for accuracy and transferability.

• adds a new Fine-tuning Guidance page covering:
  • foundation-model selection;
  • atomic reference energies (E0s);
  • method-specific optimiser settings;
  • choosing between naive fine-tuning, layer freezing, LoRA and replay;
  • validation beyond held-out energy and force errors;
• adds the Reliable fine-tuning of MACE foundation models notebook;
• adds the fine-tuning-method schematic to the guide and notebook;
• updates the main fine-tuning examples with safer defaults;
• adds the new guide and tutorial to the documentation navigation;
• documents MACE's LoRA implementation and cites the original LoRA work, ELoRA,
  and the MACE fine-tuning paper;

Tutorial

The default notebook path:

• downloads less than 10 MB of pinned supplementary data;
• runs its diagnostic plots on CPU;
• demonstrates how incorrect E0s can distort interaction-energy curves;
• provides commands for naive, layer-freeze, LoRA and pseudolabel-replay
  fine-tuning;
• shows method-specific learning-rate behaviour;
• illustrates why changing loss weights during Stage Two can destabilise
  fine-tuning;
• leaves GPU training as an explicit opt-in.

The notebook is version-controlled under docs/examples and opened in Colab
through GitHub.