Optimize QSPI drawing: async DMA, single-transaction region+pixels, larger buffers

Three targeted optimizations to close the performance gap between the
generic qspibus/displayio pipeline and the dedicated rm690b0-standalone
framebuffer driver:

1. Async DMA (QSPIBus.c): Remove blocking wait_all_transfers_done at the
   end of send_color_bytes(). DMA now runs asynchronously, allowing
   fill_area() for the next subrectangle to overlap with the previous
   DMA transfer. The wait is moved to begin_transaction(), which only
   blocks when the bus is actually needed. Python API (write_data) still
   waits synchronously for backward compatibility.

2. Single-transaction path (BusDisplay.c + bus_core.c): Add
   displayio_display_bus_send_region_commands() that sends column/row
   window commands within an existing transaction. For QSPI buses,
   _refresh_area() now combines set_region + RAMWR + pixel data into
   a single begin/end transaction, eliminating 2 of 3 per-subrectangle
   transaction cycles.

3. Buffer enlargement (BusDisplay.c): Port the QSPI-specific buffer
   boost from the rm690b0 branch (CIRCUITPY_QSPI_DISPLAY_AREA_BUFFER_SIZE
   up to 2048 uint32_t = 8KB), reducing the number of subrectangles per
   refresh and thus the total transaction overhead.

Co-authored-by: ppsx <7107135+ppsx@users.noreply.github.com>

Author: Copilot