Second review round:
- connect() awaits _pendingNativeClose before touching flserial's global
slot table. fl_close() (helper isolate) and fl_free()/fl_init() mutate
the same flserial_tab, so reconnecting during the kernel's closing_wait
could double-close the old slot or kill the freshly opened one.
- dispose() no longer re-enters disconnect() when a close is in flight;
the running disconnect owns the ordering (native close, then
subscription cancel), so dispose only chains _closeFrameController onto
the pending close.
- dispose() frees the I/O buffers through the new _releaseSerialBuffers()
on every path, not just after a successful closePort(); with _serial
nulled, an error/closed-state dispose used to leak both 8 KiB buffers.
- _handleSerialData() ignores callbacks once the status leaves `connected`:
a callback queued before flh was invalidated made readList() throw and
published a spurious transport error mid-teardown.