drivers: spi_nxp_lpspi: Fix extra byte issue on v1

The interrupt handling was not deterministic before because it relied
on "guessing" if the lpspi was a v1 type that was stalling due to design
errata. This obviously ended up being wrong in some cases. So we really
need to fix this so that it is deterministic, and my idea to do that is
to explicitly count how many words we have written to the TX fifo
throughout the whole transfer.

As a side effect of making the IRQ more deterministic, we can't support
the SPI_HOLD_ON_CS flag for v1 LPSPI anymore. It is fundamentally
impossible due to the fact that the transfer can only complete in
hardware as a result of CS deasserting. If this is absolutely a
requirement to support this flag in the future, my recommendation is to
update the driver so that it muxes the pin to a gpio before ending the
transfer, but that would kind of defeat the point of having a native CS,
at least at the end of the xfer.

Signed-off-by: Declan Snyder <declan.snyder@nxp.com>