Fix workflow deletion consistency and visibility

[physikal]

Summary

• authorize workflow deletion against the canonical workflow owner instead of the caller-supplied NIP-09 coordinate
• retire the scheduler row and kind 30620 definition event together
• allow repeated deletion to clean up orphaned workflow-definition events created by the legacy split path
• invalidate the workflow trigger cache after successful scheduler removal

Why

Workflow scheduling is driven by the database row, while Desktop and CLI discovery are driven by the kind 30620 event. The previous delete path removed only the row and trusted the pubkey embedded in the incoming coordinate. That allowed deletion events to be accepted even when their post-commit side effect could not remove the actual workflow, and left stale definitions visible.

Validation

• cargo test -p buzz-relay — 784 executed, 0 failures
• cargo clippy -p buzz-relay --all-targets -- -D warnings — passed
• just ci — all Rust, Desktop, web, analysis, and build gates passed; the run stopped on an unrelated existing mobile widget test:
  • ChannelDetailPage keeps follow mode off while a tall newest message stays visible
  • focused rerun reproduces the same failure
  • this branch changes only crates/buzz-relay/src/handlers/side_effects.rs

Buzz channel: eb4d3406-52ba-47f7-84ce-0aed2f2c084b