Refactor LPSPI with word packing, futures

[mciantyre]

Instead of using LPSPI continuous transactions, we pack the user's data into u32 words. The primitives for translating the user's data to / from the data FIFOs should help us as we consider async LPSPI drivers. And, in order to implement embedded-hal 1.0 traits, we'll need something like the dummy transmit / receive helpers, since we need to handle differing transmit / receive buffer sizes. The included unit tests don't trigger an error in Miri, and they try to simulate how we'd use the primitives in firmware. (This is an "it's not obviously wrong" test, not an "it's correct" test; help me review here.)

The commit introduces spinning futures into the LPSPI driver. By combining and spinning on these futures, we can realize in-place transfers, read-only transactions, write-only transactions, etc.

We no longer return the Busy error; we'll wait for transmit FIFO space. We also never return the NoData error, instead returning success when there's no I/O to do. Since this commit is a non-breaking change, the two errors are still available in the error enum. I'll remove them later.

I'm moving the blocking SPI example into RTIC and rewriting the driver test. The tests demonstrate overlapping writes, writes with flushes, and in-place transfers with a physical loopback. There's also tests that show how word sizes and bit orders interact. I'd appreciate if folks could test these changes in their system, since it affects how the embedded-hal implementations behave. I'm only testing this commit on a 1170EVK with the new example.