Allow unaligned input/output to SPI::transferBytes (#5709)

* Allow unaligned input/output to SPI::transferBytes

Fixes #4967

Support any alignment of input and output pointers and transfer lengths
in SPI::transferBytes.  Use 32-bit transfers and FIFO as much as
possible.

* Refactor misaligned transfer, avoid RMW to FIFO

The SPI FIFO can't properly do RMW (i.e. bytewise updates) because when
you read the FIFO you are actually reading the SPI read data, not what
was written into the write FIFO.

Refactor the transferBytes to take account of this.  For aligned input
and outputs, perform as before (but handle non-x4 sizes properly).  For
misaligned inputs, if it's unidirectional then do bytewise until the
direction data pointer is aligned and then do 32b accesses.  Fod
bidirectional and misaligned inputs, copy the output data to an aligned
buffer, do the transfer, then copy the read back data from temp aligned
buffer to the real input buffer.

* Fix comments, clean condition checks, save stack

Add more comments and adjust naming to be more informative in
transferBytes_ and *aligned_.  Save 64bytes of stack in double
misaligned case.

* Optimize misaligned transfers, reduce code size

On any misaligned input or output, always use a temp buffer.  No need
for special casing and bytewise ::transfer().  This should be faster as
bytewise ::transfer involves a significant number of IO register
accesses and setup.  Thanks to @devyte for the suggestion.

libraries/SPI/SPI.h

@@ -79,6 +79,7 @@ private:
   uint8_t pinSet;
   void writeBytes_(const uint8_t * data, uint8_t size);
   void transferBytes_(const uint8_t * out, uint8_t * in, uint8_t size);
+  void transferBytesAligned_(const uint8_t * out, uint8_t * in, uint8_t size);
   inline void setDataBits(uint16_t bits);
 };