fix(perf): use correct atomic memory operations

[cvonelm]

Use atomic_load_n and atomic_store_n. those are equivalent to the the smp_load_acquire and smp_store_release functions used in ring_buffer.h in the linux kernel.

As a proof, they produce the exact same assembly on

x86: https://godbolt.org/z/s7cMoe5vK

and

ARM: https://godbolt.org/z/6oYrKnbzv

This fixes #287

[cvonelm]

The Actual implementation further down uses smp_load_acquire() and smp_store_release()