perf(snapshot): make checkpoint publication async and deduplicated (#67)

[leinasi2014]

Summary

Finishes the remaining Issue #67 hot-path work after #88 already bounded Python snapshot traversal/context preflight.

• chunked Host checkpoint commit uses awaited fs/promises I/O and stays inside the cell settlement barrier
• the exact validated Buffer assembled from checkpoint chunks is consumed once by the Session runtime; durable publication no longer immediately rereads that snapshot synchronously
• restore-frame reads are asynchronous
• durable publication uses awaited exclusive temp/write/fsync/rename/directory-sync
• identical durable content is skipped only when both content hash and the current file fingerprint match the last validated/published generation, so an external replacement cannot be hidden by cache
• reset/rescan invalidate dedup state
• real regressions prove unchanged checkpoint content does not rewrite the durable generation and that eval settles only after the durable generation is immediately restorable
• adds pnpm bench:snapshot for repeatable 0.5/2/6 MiB ephemeral/durable measurements (p50/p95, event-loop p95, peak heap, durable bytes); benchmark has no noisy shared-runner threshold

This intentionally does not reopen #88's already-fixed bounded Python traversal and leaves startup/rescan/reset synchronous because they are not the repeated per-cell hot path.

Final Windows + Ubuntu + development-memory CI is required before merge.

Closes #67.

[leinasi2014]

Superseded by #111 which merged to main (46ce7ff).