[bdbarnett]

Building audiodsp into CircuitPython for an RP2350 board, such as the Raspberry Pi Pico 2, failed at link with undefined reference to 'lrintf' from audiodsp_graph.o. The RP2350 port doesn't link that libm function, so the README's claim that every module builds into CircuitPython 11 didn't hold there.

The graph runner called lrintf in two places: to turn a sample-and-hold rate in Hz into a whole number, and to round a clamped float sample to int16. Both now use a small static inline in audiodsp_graph.c that rounds to nearest, ties to even, which is what lrintf does in the default rounding mode. This is the same approach audiodsp_trig took for fmod on the RP2040 (#210, #225).

Nothing changes on other builds. The helper agrees with lrintf for every float up to 2^24 in magnitude, both signs (about 2.5 billion values, checked exhaustively), which covers any sample rate and any int16 sample. No other lrint or fmod call remains in src/shared or src/circuitpython_spike.

How it was checked, with CircuitPython main at fb98f214c and arm-gnu-toolchain 14.2.rel1:

• raspberry_pi_pico2: before the change, the link fails with five undefined reference to 'lrintf'. After it, the link is clean and the .uf2 is 1,995,776 bytes (997,804 bytes of flash used).
• pimoroni_pico_plus2w: links clean with every audiodsp module, .uf2 of 3,079,680 bytes (1,539,772 bytes of flash used).
• The graph tests pass: test_cpython_audiograph.py and test_cpython_audiograph_nodes.py (25 passed), graph_probe.py on a unix MicroPython build (GRAPH PASS), and verify_graph.py comparing CPython with that MicroPython at a tolerance of 0 counts (0 samples differ). Each planted fault in both gates still fails.

Fixes #311