fix: harden multithreading synchronization and v2 docs

[csheldrick]

Follow-up to #30 focused on threading correctness, portability, and a few v2 regressions found after the multithreading rewrite.

What changed

• Restore opt-in stb-style POLESITTER_IMPLEMENTATION semantics.
• Synchronize thread-pool queue/state access under the existing spinlock instead of relying on volatile.
• Avoid requiring AVX intrinsics just to issue an x86 spin-wait hint.
• Fix the ST-vs-MT equivalence test so force mismatches actually fail the test.
• Update the README quickstart to match the v2 API, including:
  • PS_MULTITHREADING
  • max_particles
  • configurable theta
  • thrd_cnt
  • ps_destroy
  • removal of the now-internal arena reset call

Why

The multithreading refactor introduced a few cross-cutting regressions while changing the execution architecture.

In particular, the thread pool was reading cnt, active_jobs, and shutdown_flag concurrently outside the spinlock and relying on volatile. In C, volatile does not provide atomicity or inter-thread synchronization, so those accesses constitute a data race.

The updated implementation keeps the existing spinlock architecture but ensures shared pool state is inspected and modified while holding the lock.

The ST-vs-MT comparison test also printed force mismatches without incrementing its mismatch counter, allowing that check to report success despite differing forces.

Finally, the v2 README example still reflected the previous API and would not correctly initialize the new radix/threading state.

Validation

The final branch passes:

• clang-format
• clang-tidy
• Ubuntu build/tests
• macOS/ARM build/tests
• Windows build/tests
• threaded ST-vs-MT equivalence validation
• existing direct-reference numerical accuracy tests

This intentionally leaves the overall threading architecture unchanged and limits the patch to correctness, portability, and documentation fixes.

[nihiL7331]

Looks good. Merging now.