drivers: serial: nrf uarte: avoid dropping RX chars/overruns

In some cases (eg at high baud rate, no HW flow control, and when BLE
radio/ints running) data could be lost between when enough characters
have been RX'd to fill the DMA buffer and when the ENDRX event was
fired, where the the STARTRX task is invoked to start filling the next
buffer (which is set up earlier, but I think will not be filled until
STARTRX).
To fix this, the SHORT is enabled between ENDRX and STARTRX whenever the
'next' buffer is available, so that STARTRX is invoked automatically and
subsequent chars go into the next buffer via EasyDMA.
To make this work properly, uarte_nrfx_isr_async() now handles the ENDRX
event _before_ the STARTRX event.

There was also an issue in rx_timeout() where the received character
count (rx_total_byte_count) could be incremented greater than the actual
buffer size. This arises from rx_total_byte_count value coming from the
counting the RXDRDY events (either by PPI/timer counter or counting the
RXDRDY ints themselves) and so if chars are received in the rx_timeout()
(or before ENDRX is handled) the rx_timeout() could increment rx_offset
past the length of the buffer. This could result the remaining 'len'
being calculated incorrectly (an underflow due to unsigned - signed ,
where signed > unsigned).
To fix this, we now store the lengths of the buffers and don't invoke
the UART_RX_RDY callback when the buffers are full; its handled by
ENDRX.
(Also note that the buffer size should be available via the RXD.MAXCNT
register on the nrf, but this register is not exposed through the nrfx
HAL and is also double buffered, so it seemed clearer to just track the
buffer lengths explicitly here in the driver).

Signed-off-by: Marc Reilly <marc@cpdesign.com.au>

for fixup

Signed-off-by: Marc Reilly <marc@cpdesign.com.au>