Fix F-07: pre-compute real Argon2 dummy hash via sync.Once

- auth/auth.go: add DummyHash() which uses sync.Once to compute
  HashPassword("dummy-password-for-timing-only", DefaultArgonParams())
  on first call; subsequent calls return the cached PHC string;
  add sync to imports
- auth/auth_test.go: TestDummyHashIsValidPHC verifies the hash
  parses and verifies correctly; TestDummyHashIsCached verifies
  sync.Once behaviour; TestDummyHashMatchesDefaultParams verifies
  embedded m/t/p match DefaultArgonParams()
- server/server.go, grpcserver/auth.go, ui/ui.go: replace five
  hardcoded PHC strings with auth.DummyHash() calls
- AUDIT.md: mark F-07 as fixed
Security: the previous hardcoded hash used a 6-byte salt and
  6-byte output ("testsalt"/"testhash" in base64), which Argon2id
  verifies faster than a real 16-byte-salt / 32-byte-output hash.
  This timing gap was measurable and could aid user enumeration.
  auth.DummyHash() uses identical parameters and full-length salt
  and output, so dummy verification timing matches real timing
  exactly, regardless of future parameter changes.

AUDIT.md

@@ -227,7 +227,7 @@ The REST `handleTokenIssue` and gRPC `IssueServiceToken` both revoke the existin
 | Yes | F-11 | MEDIUM | Missing security headers on UI responses | Small |
 | No  | F-05 | LOW | No `nbf` claim in issued JWTs | Trivial |
 | No  | F-06 | LOW | `HasRole` uses non-constant-time comparison | Trivial |
-| No  | F-07 | LOW | Dummy Argon2 hash timing mismatch | Small |
+| Yes | F-07 | LOW | Dummy Argon2 hash timing mismatch | Small |
 | No  | F-08 | LOW | No account lockout after repeated failures | Medium |
 | No  | F-09 | LOW | `synchronous=NORMAL` risks audit data loss | Trivial |
 | No  | F-10 | LOW | No maximum token expiry validation | Small |