Improve clone performance

The previous MR in #2 was on the right track — caching the UUID delta maps so we don’t rebuild them for every paragraph save is a big win on large nodes.

However, the original change used request-wide static caches. This caused cross-field data leakage when cloning nodes that have more than one layout_paragraphs reference field (e.g., a “content” field and a “modals” field). In those cases, the maps built for the first field were reused for subsequent fields, breaking the parent/child structure in the second field.

This MR in #5 keeps the performance improvement while avoiding that regression by:
- Introducing a LayoutParagraphsParentUuidCloneMapper service to encapsulate the mapping logic.
- Caching maps per clone context — keyed by (source host, destination host, field name) — so maps are only reused when it’s actually safe.
- Falling back to an unscoped (*) map only if the scoped map does not contain the required UUID, preserving compatibility for edge cases.
- This approach still builds the source and destination maps only once per context, so cloning large, complex nodes remains fast, while ensuring multiple layout_paragraphs fields on the same node clone correctly.