Fix TX underruns and implement the shared dmdrvi audio stream contract

[JohnAmadis]

1. Wake SAI writers from the DMA interrupt instead of sleeping a tick (372b0fd)

Problem

stm32_sai_dma_write/read/flush waited for queue space with Dmod_ThreadSleep(1). dmod-boot runs FreeRTOS at a 100 Hz tick, so that sleep lasts about 10 ms. That is longer than the whole TX path at 48 kHz: a 256-frame queue plus two 64-frame DMA halves is about 8 ms. Every full queue therefore ended in about 4 ms of silence: choppy, crackling playback and thousands of tx_underruns.

Fix

• The DMA callback (half, complete, error) and stm32_sai_dma_stop post a binary dmosi semaphore.
• wait_queue waits on that semaphore for at most 10 ms, instead of sleeping.
• stop takes any pending wake before posting.

Verification (Renode, 8 s melody through /dev/wm8994)

	tx_underruns	output length
before	4887	14.2 s
after	4 (drain tail)	7.94 s

2. Shared dmdrvi audio stream contract instead of DMSAI_IOCTL_* (9c7f004, e90310f)

The node implements the DMDRVI_IOCTL_AUDIO_* stream commands from choco-technologies/dmdrvi#28. A codec node implements the same contract, so an application plays the same way on the codec and on the bare SAI.

• AUDIO_GET_STREAM returns dmdrvi_audio_stream_t: the buffer layout (rate, format, channels = active slots) and the link (framing, clock role, slots).
• AUDIO_CONFIGURE returns 0 only when rate, channels and width match the ini format; otherwise -ENOTSUP.
• AUDIO_START succeeds when the stream is already running (it used to return -EBUSY).
• AUDIO_STOP, AUDIO_GET_STATUS, AUDIO_SET/GET_TIMEOUT and AUDIO_DRAIN behave as before.
• AUDIO_GET_INFO, SET_VOLUME and SET_MUTE return -ENOTSUP, since the SAI has no analog stage. This also keeps dmdevfs from reporting the SAI as an audio device.
• dmsai_ioctl.h and DMSAI_IOCTL_* are removed. The port API keeps dmsai_config_t / dmsai_status_t, and static asserts check that format and clock role values match dmdrvi.
• README, docs/api-reference.md and the driver board test are updated. The board test now also checks CONFIGURE matching and mismatching, the -ENOTSUP volume control, and the idempotent START.

Requires dmdrvi 2.7. CI cannot fetch it until choco-technologies/dmdrvi#28 is merged and released. Merge this together with choco-technologies/dmwm8994#3.

Verified locally: built for cortex-m7 against the new header and ran in Renode with a generic player that uses only the shared commands.

• Through /dev/wm8994: all 32 notes are correct.
• On /dev/dmsai1 directly: the stream runs without underruns (drain tail only). It is silent there because nothing configures the codec on that path.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Nr1HhezRY951yBxxTm1b65