Make learner backends optional

[ielbadisy]

Summary

• Keep the core funcml package lightweight by moving learner backend engines from hard imports to suggested packages.
• Add a central backend availability check in fit() so model-specific errors remain clear.
• Guard SHAP plotting conversion with the same package check helper.

Impact

• Packages that import funcml, including mimar, still get the funcml API automatically.
• Heavy engines such as torch, lightgbm, xgboost, dbarts, shapviz, and other learners are only needed when the corresponding learner is used.

Validation

• funcml tests: 251 passed, 0 warnings.
• R CMD check --no-manual --no-build-vignettes funcml-package: no errors; remaining warnings/notes are existing source-tree packaging artifacts.
• Temp source install of funcml and mimar succeeded; mimar default imputation smoke test passed with the temp-installed funcml.