[fiber] Use synchronization primitives without scheduler

[salkinium]

Follow-up to the missing volatile fixes in #1386. Those were only a symptom: without modm:processing:fiber the yield() function is an empty inline stub, so the compiler assumes that nothing changes while polling and happily hoists the load of the flag out of the loop:

bool done; // set in an interrupt
while (not done) modm::this_fiber::yield();
// GCC -O2: the whole loop is gone, whoops

So instead of hunting down every non-volatile flag in every driver, the stub now tells the compiler that memory changed, just like a real context switch does.

• yield() stub contains an empty asm with a memory clobber. Emits no code, but the flag gets reloaded on every iteration.
• mutex, shared_mutex, semaphore, latch, stop_token and condition_variable are now also available without the scheduler. They only need yield() and get_id(), so they work just fine for synchronizing the main context with interrupts.
• barrier is not available without the scheduler, since it needs multiple fibers to make any progress.
• get_id() stub returns the interrupt number inside interrupts on Cortex-M, same as the scheduler. It returned 0 everywhere before, so the recursive mutex could not tell the main context apart from an interrupt.
• modm:architecture:fiber depends on modm:architecture:atomic for the recursive mutex.
• Docs: without the scheduler a blocking lock(), acquire() or wait() waits in place and only returns if an interrupt releases it.
• New modm-test:test:architecture:fiber unit test that uses the primitives from outside of a fiber, so it runs with and without the scheduler.
• New hosted-no-fiber unit test configuration and CI step for the latter.

Tests:

• Hosted unit tests pass on macOS arm64 with and without the scheduler.
• Cortex-M without the scheduler compiles and links on NUCLEO-F103RB.
• AVR and STM32 unit tests compile with unchanged sizes for the configurations I tried.
• run-hosted-linux-no-fiber in CI, only ran the macOS equivalent locally.
• Hardware test with an actual interrupt releasing a semaphore.

This is only the first half: the plan is to then replace the hand-rolled locks in modm (SpiLock, ResumableMutex, the AMNB node, the STM32 SPI DMA flags) with these primitives in a separate PR.