dcd/samx7x: always send deferred ZLPs from the TXINI handler

The immediate path raced the hardware: the last busy bank could free between the TXINI acknowledge and
the NBUSYBK read, and the fresh TXINI then reported the ZLP complete while it still sat in the bank.
NBUSYBK also could not tell a free current bank from a busy one on dual-bank endpoints.
The ZLP is now always validated from the TXINI handler, where the bank is guaranteed free. If a bank
is writable already, TXINI is raised by software the way the ASF driver does, so there is one path
and the completion always follows the transmission