UAC: take each packet in the USB interrupt, so a busy interpreter no longer costs packets (#39)

[bdbarnett]

Fixes #39 on the ESP32-P4. With the interpreter busy, the P4 sound card now gets the whole stream, and the steady tick under the music is gone.

Needs micropython-pydevices#21; merge that first. It gives TinyUSB 0.21 to the P4 only. The S2 and S3 stay on 0.18 for now (#48). There this change builds and behaves as before, so an S3 sound card still ticks under load. That remainder lives in #48.

What changed

• The packet hook runs in the USB interrupt. On TinyUSB 0.19+ it's tud_audio_rx_done_isr(), and TinyUSB has already re-armed the endpoint by the time it runs. On older TinyUSB it's still tud_audio_rx_done_post_read_cb(). Either way it does only ISR-safe work: two counters, a FIFO level read and a task notification. No MicroPython, no FIFO reads, nothing that blocks.
• usbif_pump_notify() checks where it's called from. From the ISR it uses vTaskNotifyGiveFromISR + portYIELD_FROM_ISR; from a task (rate change, stop) it uses xTaskNotifyGive.
• The no-pump "shed to half" is gone. It read the FIFO from the callback, and an ISR mustn't do that while Python may be reading too. Feedback is disabled, so a full FIFO no longer stalls the host, and TinyUSB's FIFO overwrites its oldest audio. uac_stats() still counts overflows (FIFO above three-quarters with no pump).
• High speed goes back to bInterval 1 (a packet every microframe) on 0.19+. It has to. 0.21's DWC2 driver arms an isochronous OUT endpoint for the microframe after the current one, so at bInterval 4 the host's packet lands on the other frame parity and is dropped. The P4 got no packets at all at 4. Re-arming in the ISR is what makes 125 µs safe now. On older TinyUSB it stays 4 (Sound card over high speed: the pump receives 80 % of the stream, at every rate #37).
• TUD_AUDIO_EP_SIZE took a speed argument in 0.20; a small wrapper keeps both versions building.

Evidence

This was on the P4 panel, using tests/hardware/uac_stream_check.py plus uac_play_tone.py from Windows WASAPI, with a 440 Hz sine, 10 s windows and at least 80 s per run. The board volume was 30 %.

build	load	host	delivered	pitch at I2S
main (TinyUSB 0.18)	busy (STRESS=1)	48 kHz shared	90.6 to 95.5 %, every window below 99.5	441 to 576 Hz, unstable
this PR + 0.21	busy	48 kHz shared	100.00 % (8,000.1 to 8,000.4 packets/s)	440.0, 439.8
this PR + 0.21	idle	48 kHz shared	99.92 to 100.00 %	440.2; 438.5 and 446.7 in shared mode
this PR + 0.21	busy	44.1 kHz exclusive	99.89 to 100.00 %	440.0 ×3
this PR + 0.21	busy, 48 kHz wire as in soundcard.py	48 kHz exclusive	100.00 % after the stream-start window (99.46)	440.0 ×3

The checker failed on the old build (every window below the 99.5 % gate) and passes on the new one. At 44.1 kHz the pump still times out 7 or 8 times per 10 s, which is #40 (no feedback), unchanged.

Also on the 0.21 image:

• All 63 costumes enumerate and mount. The sweep waits to see each one detach and come back.
• The CDC REPL works on the same cable while the card streams: 8,000.5 packets/s during an exec, and 7,999.8 during a 200-line print.
• MSC works: a FAT RAM disk read back on Windows.
• The change builds against the old TinyUSB too.

Both TinyUSB versions, one source (2026-09-25)

Both boards ran images from this branch (b7ee671; the later commit only touches the README) on micropython-pydevices#21's per-target overlay, with a quiet 440 Hz tone over WASAPI, 48 kHz shared.

board	TinyUSB	hook linked	load	delivered
P4 panel (high speed)	0.21 (v0.21.0.1-micropython1)	tud_audio_rx_done_isr	busy	100.00 %, 8,000.2 to 8,000.3 packets/s, 8 windows after stream start; pitch 440.0
T-Embed (S3, full speed)	0.18 pin	tud_audio_rx_done_post_read_cb	idle	99.91 to 99.93 % (999.1 to 999.3 of 1,000 packets/s)
T-Embed	0.18 pin	tud_audio_rx_done_post_read_cb	busy	87.4 to 93.9 %, #39 as expected on 0.18
T-Embed, an experiment for #48	v0.21.0.1-micropython1	tud_audio_rx_done_isr	busy	100.00 %, 1,000.0 packets/s, 9 windows

The version guards pick the right callback, TUD_AUDIO_EP_SIZE form and high-speed interval on each: both builds compile, and the link maps name the hook shown. On the T-Embed the pitch at I2S read 440.0 in a mono run; the checker's tap only reads a mono sink. On 0.18 under load the T-Embed's CDC console also went silent to the host, so its busy runs logged to a file on the board.