Move various imports inside methods/functions to load them lazily when initializing the app

[Irozuku]

Summary

This pull request refactors import statements across multiple component modules to use lazy imports for third-party dependencies. Heavy libraries are now imported only when the relevant functionality is invoked, reducing startup overhead and improving startup initialization time when the application is built and distributed as an .exe.

---

Type of Change

Check all that apply like this [x]:

• Backend change
• Frontend change
• CI / Workflow change
• Build / Packaging change
• Bug fix
• Documentation

---

Changes (by file)

• Component modules (HuggingFace, imbalanced-learn, scikit-learn):
  Moved imports of pyarrow, torch, datasets, transformers, imblearn, sklearn, numpy, and pandas from the module level into methods or constructors to enable lazy loading.
• Method implementations (get_output_type, _load_model, _process_batch, transform, fit):
  Added local imports so dependencies are loaded only when the corresponding functionality is used.
• Type annotations and internal state:
  Updated type hints and attribute initialization to remove direct references to lazily imported libraries.

---

Testing (optional)

• Verify that affected components still function correctly when executed.
• Confirm that component modules can be imported without requiring optional dependencies.
• Check that application startup is faster when running the packaged .exe.

---

Notes (optional)

This change is especially beneficial for packaged distributions, as it reduces import-time overhead and avoids bundling or loading unused dependencies during application startup.