STM32F4: SPI::abort_transfer() causes incorrect data on the next transaction

[multiplemonomials]

Seeing an issue that's showing up in the SPI unit tests, specifically async_queue_and_abort() from the spi-basic test:

After the gap, the bytes should be "01 02 00 00", but instead the first two bytes are zeroes and the next two nonzero. Maybe it's somehow swapping the order of the first 4 bytes?

This issue seems to affect both interrupt and DMA SPI, so I don't think it's a DMA-specific problem.

I observed this on NUCLEO_F429ZI, but I wouldn't be surprised if it affects all STMicro targets with SPI IP v1.