Parallelize checkpoint archive construction for large dirty worktrees

[jbellis]

Problem

Importing or checkpointing a session snapshots each repository serially. On a dirty brokkbench session import, Git snapshot/archive construction remained CPU-bound for minutes before native resume could start.

This is a worst-case input, but it makes archive creation dominate session import and close latency.

Observation

The importer should continue to use git ls-files --others --exclude-standard: repository-local ignore rules select valid inputs. Hel should not blanket-skip arbitrary directory types. The follow-up is about throughput after Git has selected the input set.

Desired direction

Parallelize independent archive work where safe (repository collection, payload digest/compression, and possibly archive writing), while preserving:

• atomic archive creation and verification;
• bounded memory and disk use;
• deterministic manifests;
• the current rejection of unsafe or unsupported file types; and
• no writes to the user worktree.