feat: auto-follow basic group → supergroup target migration

[awdr74100]

Problem

When a forward target is a basic group that later gets upgraded to a
supergroup, its id changes (e.g. -4187363166 → -100…). Telegram then
rejects every forward to the stale id with CHANNEL_INVALID /
PEER_ID_INVALID. Before this change that surfaced as a recurring
Forward failed error on every source message, and the only fix was to
manually re-run group edit.

What changed

The forwarder now recovers from this automatically. When a forward fails
with an invalid-target error, it probes the target with getChat:

• Migrated → write the new supergroup id back to the config and retry
  the forward transparently on the new id (logged once, info level).
• Still healthy / not migrated → re-raise the original error; the
  target was not the cause, so a good target is never wrongly retired.
• Unreachable → retire it for this session (SkippedForward) and skip
  it silently after a single warning, instead of spamming the log on every
  source message.

FloodWait, deleted-message, and transient network errors are unaffected —
they still flow to the queue's existing retry logic.

Details

• filter.ts: add replacePeer (swap a stored peer id, preserving order
  and dropping duplicates).
• queue.ts: add SkippedForward sentinel so a deliberate skip is
  recorded as skipped (warning) rather than failed (error).
• forwarder.ts: forwardOne / probeTarget migration handling,
  injectable persistConfig (defaults to saving the real config),
  per-session deadTargets set, and isInvalidTargetError.

Tests

• replacePeer: swap / no-op / dedupe / empty.
• isInvalidTargetError: matches the invalid-peer codes by errorMessage
  and message string, and does not swallow FloodWait or deleted-message
  errors.

pnpm typecheck, pnpm lint, pnpm test (61 passed), pnpm build, and
pnpm knip all pass.

Note: the forwardOne migration orchestration and the SkippedForward
queue branch are not directly unit-tested (they require driving a real
Dispatcher + RPC, which this repo deliberately does not mock). The pure
logic they rely on is covered.