fix: recover from a card swap without a server restart (v3.5.2)

Equip was broken after a card swap: auto-equip (and manual Equip) failed with
"Failed to transmit with protocol T0. Card was removed. (0x80100069)" until
the server was restarted.

Chain (v3.1.2 regression):
- _handle_card_disconnect() cleared server.card/scc but not pySim's
  app.card/app.rs/app.lchan, so the removed card - and with it the old PC/SC
  link and its exclusive card handle - stayed referenced; handlers using
  app.rs (e.g. the PWA's /api/tree poll) kept transmitting over the dead card.
- those errors carry hresult=0x80100069 (SCARD_W_REMOVED_CARD), but
  _is_pcsc_error() treated any PC/SC error as a dead service and set
  _TRANSPORT_STALE; the next auto-equip then called _ensure_transport(),
  which built a second PcscSimLink while the old one was still connected -
  the new link inherited the removed card's handle and failed on its first
  APDU.  No retry existed, so every later equip repeated the failure.

Fixes:
- _is_transport_fatal(): only service/context hresults rebuild the transport
  (E_NO_SERVICE, E_SERVICE_STOPPED, E_NO_READERS_AVAILABLE, E_INVALID_HANDLE,
  ...); card-level states (W_REMOVED_CARD, E_NO_SMARTCARD, W_RESET_CARD,
  W_UNRESPONSIVE_CARD, W_UNPOWERED_CARD) reconnect on the existing link.
  All 8 disconnect call sites pass the new verdict.
- _clear_app_card_state(): unequip through pySim's own equip(None, None)
  before clearing app.card/app.rs/app.lchan.  Nulling them alone would make
  the next equip abort with "CommandSet ... is already installed"
  (PysimApp.equip() unregisters the previous profile's command sets from
  self.rs), which left /api/tree broken after a swap.  A failed auto-equip
  attempt unequips the half-initialized shell as well.  Handlers now answer
  "no card" instead of transmitting over the dead card, and the old link
  becomes collectable.
- _ensure_transport(): disconnects the old link before building the new one
  (restoring it if the factory fails) - the rebuild path is now safe for the
  real pcscd-restart case.
- auto-equip: _auto_equip_attempt()/_auto_equip_attempts() retry up to 3
  times, 1 s apart (fatal failures mark the transport so the retry rebuilds);
  the watchdog re-arms it (_auto_equip_rearm) every 5 s while a card is
  present and the session is down, with exponential backoff to 60 s for a
  card that cannot be initialized at all.
- fastinit.init_card_fast(): also retries once after a PC/SC link error
  (CardConnectionException/NoCardException), not only after SW mismatches.

Tests: transport-fatal classification, app-state clearing, shell unequip,
old-link release, auto-equip retry/backoff/re-arm, fastinit link retry.
547 frontend / 410 Python green; version 3.5.2; sw cache simple-v254.

tests/test_fastinit.py

@@ -214,6 +214,30 @@ class TestInitCardFastRetry(unittest.TestCase):
         self.assertEqual((rs, card), ('rs', 'card'))
 
 
+class TestInitCardFastLinkRetry(unittest.TestCase):
+    def test_link_error_retries_after_reconnect(self):
+        from smartcard.Exceptions import CardConnectionException
+        calls = {'once': 0, 'disconnects': 0}
+
+        class FakeLink:
+            def disconnect(self):
+                calls['disconnects'] += 1
+
+        def once(sl, skip, wait):
+            calls['once'] += 1
+            if calls['once'] == 1:
+                raise CardConnectionException(
+                    'Failed to transmit with protocol T0. Card was removed.',
+                    hresult=0x80100069)
+            return ('rs', 'card')
+
+        with mock.patch.object(fastinit, '_init_card_once', side_effect=once):
+            rs, card = fastinit.init_card_fast(FakeLink(), wait=True)
+        self.assertEqual(calls['once'], 2)
+        self.assertEqual(calls['disconnects'], 1)
+        self.assertEqual((rs, card), ('rs', 'card'))
+
+
 class TestDoEquipFastFailure(unittest.TestCase):
     def test_failed_equip_keeps_previous_state(self):
         calls = []