turbo: add CIRCUITPY_LOAD_NATIVE to load host-compiled native .mpy

[mikeysklar]

Second of five PRs for turbo. #11331 backported the native .mpy loader but left it compiled out. This one adds the CIRCUITPY_LOAD_NATIVE build flag that turns it on, and enables it on five boards. Depends on #11331.

What

CIRCUITPY_LOAD_NATIVE=1 loads native and viper .mpy built on the host by mpy-cross, without compiling in the on-board emitter. On-board compiling of @native and @viper is untouched and still lives behind CIRCUITPY_ENABLE_MPY_NATIVE.

Why

The emitter costs 20 to 50 KB. On the Metro M4 AirLift Lite it does not fit at all: CIRCUITPY_ENABLE_MPY_NATIVE=1 fails to link, region FLASH_FIRMWARE overflowed by 13104 bytes (12.8 KB). The loader alone:

Board	Loader cost	Free after
Metro M4 AirLift Lite	+1.9 KB	6.7 KB
Metro RP2040	+2.9 KB	56.8 KB
Metro RP2350	+2.7 KB	120.5 KB
Feather nRF52840 Express	+2.4 KB	136.8 KB
Feather STM32F405 Express	+2.4 KB	365.3 KB

Tested on hardware

All five flashed with 10.3.0-52-gcee7c17c08 built from this branch, then driven over the serial REPL.

Board	_mpy arch	bytecode	viper .mpy
Metro M4 AirLift Lite	armv7emsp	152 ms	10 ms
Metro RP2040	armv6m	176 ms	9 ms
Metro RP2350	armv7emsp	104 ms	5 ms
Feather nRF52840 Express	armv7emsp	351 ms	18 ms
Feather STM32F405 Express	armv7emsp	158 ms	9 ms

Both variants returned checksum 3264000 on every board, so the compiled module computes what the source does. On all five, a wrong-arch (xtensawin) .mpy raises ValueError: incompatible .mpy arch, and @micropython.viper in source raises SyntaxError: invalid micropython decorator.

Not tested: Cortex-M7, non-English translations.

How I tested it

mpy-cross -march=armv7emsp -o vsum.mpy vsum.py
mpy-cross -march=xtensawin -o vsum_wrong.mpy vsum.py

# vsum.py; vsum_bc.py is the same file without the decorator
import micropython

@micropython.viper
def sum_bytes(buf: ptr8, n: int) -> int:
    total = 0
    for i in range(n):
        total += buf[i]
    return total

>>> import time, vsum, vsum_bc
>>> b = bytes(range(256)) * 100        # 25,600 B, fits every board here
>>> for m in (vsum_bc, vsum):
...     t = time.monotonic_ns(); s = m.sum_bytes(b, len(b)); print(m.__name__, s, (time.monotonic_ns() - t) // 1000000, "ms")
vsum_bc 3264000 104 ms
vsum 3264000 5 ms

Scope

Metro M0 Express is left out, and it is closer than it looks. Stock leaves 1,660 bytes free and the loader needs about 1,868, so CIRCUITPY_LOAD_NATIVE=1 fails to link by 208 bytes:

region `FLASH_FIRMWARE' overflowed by 208 bytes

Dropping safemode.py clears it, but that is a poor trade for a shipping board and not something I want to propose. One maintainer thinks roughly 700 bytes could be recovered on the SAMD21, which would cover it three times over. Suggestions for where those 208 bytes should come from are very welcome, and I am happy to build and measure any candidate. If a way is found, the board is a one-line addition to this PR.

AI assistance

Claude Code wrote the config lines and test script. I chose the boards, ran each one, checked the numbers.