speed up large-project saves and reloads

[mrhyde]

On a 2000-task project the end-to-end mutation latency was ~48s. After this pass it's ~1s.

• only rewrite changed task files on save (dirty-set tracking per project)
• suppress self-write modify/delete events instead of letting them trigger a 300ms debounced reload
• walk the projects/tasks folders directly instead of getMarkdownFiles().filter(...) over the whole vault
• pre-group children by parentId in the table renderer (O(N) instead of O(N²))
• reuse the childIds set in loadTasksFromFolder self-heal pass (O(N) instead of O(N²))

Tests cover dirty-set efficiency per mutator, self-write marker semantics (set, single-use consume, stale-after-window), title-rename child rewrites, and the concurrent-save race.

[mrhyde]

add per-project task index, making findTask/findParentId O(1) so bulk mutators don't tree-walk per id