drivers: dma: dma_mcux_lpc: fix multiple bugs with continuous mode

There are multiple bugs related to continuous/circular mode.
Continuous/circular mode is where the DMA runs continuously
until the stop API is called, instead of auto-stopping on
completion on a single transfer. After a stop, the DMA
can then be reconfigured/restarted.

1. Fix bug where stop didn't actually stop. This can cause memory
   corruption if the user thought the DMA stopped and repurposed
   the dest memory, but in fact the DMA is still writing to it.
   The bug was due the incorrect usage of the DMA controller busy
   state. The DMA controller is busy only when a transfer is
   actively in progress, but the driver needed to stop even if
   the transfer is not active but is only enabled (and may become
   active on a subsequent trigger event). Change so that data->busy
   doesn't use the DMA controller busy state but tracks the enable
   state. Also, to make it doubly safe, make stop function always stop
   regardless of data->busy state because it is alway safe/correct
   to do so.

2. Fix race condition where a stop request from another ISR might race
   with a DMA completion interrupt, and the DMA completion callback
   gets invoked after the DMA has already been stopped. The fix
   is to unregister the callback with the sdk DMA driver, so the
   ISR still runs and clear the interrupt without invoking the
   callback. There is potentially still a race if the interrupt
   is restarted before the ISR fires, so the callback might be
   called too early. However, the Zephyr DMA driver doesn't
   have the channel level details that the SDK driver does and
   it cannot clear just the channel interrupt.

Also a couple of general fixes/improvements:

a. Use interrupt B for end of transfer (single transfer or end
   of block list). Use interrupt A for interrupts of a block
   in the middle of a transfer or for continuous/circular transfers.
   This fixes the dma callback so it can properly report
   DMA_STATUS_BLOCK vs DMA_STATUS_COMPLETE.

b. Reorder some fields in struct channel_data to pack a little
   better in memory

Signed-off-by: Mike J. Chen <mjchen@google.com>