Avoid AtomicU64 if possible, fix race condition

[bugadani]

AtomicU64 is implemented with a critical section on the ESP32{|-S3}. Multi-core aware critical sections are rather expensive, so let's see if this PR can reduce that cost.

On the S2 this is less of an issue. It is a single-core SoC, so a critical section is much smaller in number of instructions. Still I hope there is value in:

• consistency
• fewer multiplications
• 4 bytes less RAM used

Secondly, the algorithm used to calculate the time was previously incorrect. If the cycle counter just overflowed, it was previously possible to add together the old global overflow counter with the now-overflowed cycle counter, resulting in a smaller timestamp than expected.

[MabezDev]

Looks good from my perspective, makes more sense to count overflows than CPU ticks.

[bjoernQ]

LGTM 👍