Persist audit actions to a tamper-evident hash-chained store

[pavsoss]

Summary

Part 1 of 2 for #1023. Audited actions were only written as free-text log
lines — editable after the fact and impossible to query. This adds a durable
audit store backed by SQLite (no new dependencies) that links every record
into a SHA-256 hash chain: each record commits to the hash of the one before
it, so editing or deleting any record invalidates every hash that follows. A
verify_chain routine recomputes the chain and reports whether the trail is
intact.

The existing audit_log decorator now persists a structured record — actor,
action, resource, correlation id, resulting status and a UTC timestamp — in
addition to the log line it already writes. Persistence is fail-soft: a store
failure is logged and swallowed so it can never turn a successful request into
an error. The database location is configurable via AUDIT_DB_PATH.

Parts

• Part 1 of 2 (this PR) — the tamper-evident store and decorator wiring.
  Targets the default branch.
• Part 2 of 2 ([Feature]: Tamper-evident, queryable audit trail for the Flask ML API #1023) — admin query endpoint, retention pruning, and docs,
  stacked on top of this.

Test plan

Regression coverage for appending records and confirming the chain verifies,
plus proof that a field edit, a status edit and an interior row deletion are
each detected as tampering.

Refs #1023

[vercel]

@pavsoss is attempting to deploy a commit to the Aditya Sharma's projects Team on Vercel.

A member of the Team first needs to authorize it.