Fix long-running backup resource handling

[itsDNNS]

Summary

• write scheduled and manual backup archives to disk-backed temporary files instead of buffering complete archives in memory
• publish scheduled backups atomically and clean temporary files after failures or completed downloads
• retain long-running collector futures until completion and prevent duplicate submissions, including during collector hot-swaps
• add regression coverage for backup lifecycle and delayed collector completion

Verification

• python -m pytest tests --ignore=tests/e2e -q: 3002 passed, 1 skipped
• focused backup and collector tests: 84 passed
• 64 MiB incompressible archive smoke test: peak RSS reduced from 156.0 MiB to 27.7 MiB with the same archive size
• git diff --check and Python compilation passed

Fixes #738