zephyr-cp: build ulab, and let boards opt out

[mikeysklar]

What

Builds ulab on zephyr-cp and adds CIRCUITPY_ULAB so boards that cannot spare the flash can turn it off. On by default; two boards opt out.

Why

zephyr-cp builds without py/circuitpy_mpconfig.mk, so the ulab wiring every other port gets from py/py.mk does not exist here and ulab could not be built at all. This adds the flags and sources the same way py.mk does.

It costs roughly 90 KB, so the two boards without room set CIRCUITPY_ULAB = false in their circuitpython.toml:

• nordic_nrf7002dk is at 99.73% of flash, under 3 KB free.
• st_nucleo_n657x0_q has no internal user flash and executes from RAM, which is already 93% full. CI caught this one; ulab overflowed it by 56520 bytes.

Every other board has at least 135 KB of headroom in whichever region holds code.

Hardware tested

Build only, which is the whole surface of this change. On silabs_siwx917_dk2605a:

build	flash
main	799136 B, 76.21%
this PR, ulab off	799136 B, byte-identical to baseline
this PR, ulab on	890168 B, 84.89%, +91032 B, 163 ulab sources

Both opt-out boards were built locally to confirm they link, and the full CI matrix is green.

Not tested: importing ulab on hardware.

Scope

Build plumbing and two board opt-outs. No docs change.

AI assistance

Written with Claude Code. I ran the builds and compared the figures myself.

[tannewt]

Please do include some changes in the circuitpython.tomls to show how it looks.

[mikeysklar]

Defaults on now, and bool() is gone, TOML gives a real bool.

Added the opt-out to nordic_nrf7002dk to show how it looks. It is the only board that needs one. Flash headroom before ulab, from the last green run:

nordic_nrf7002dk      2584 B      99.73% used
nordic_nrf54l15dk   135524 B      81.41% used
everything else     more

ulab costs about 91 KB, measured on siwx917_dk2605a, 799136 to 890168 bytes. That is headroom from the logs rather than a build with ulab on, so CI on this push is the real check.

[tannewt]

Thank you!