ForkUnion v3

This PR introduces Rayon-style parallel iterators for Rust with explicit pool control, adds cache-aligned scratch space primitives, and delivers significant performance improvements across the board.

Rayon-style Parallel Iterator API for Rust

Added ergonomic parallel iterator support with explicit control over thread pools and scheduling:

use fork_union::prelude::*;

let mut pool = fu::spawn(4);
let data: Vec<usize> = (0..1000).collect();

(&data[..])
    .into_par_iter()
    .filter(|&x| x % 2 == 0)
    .map(|x| x * x)
    .with_pool(&mut pool)
    .for_each(|value| println!("Squared even: {}", value));

Key differences from Rayon:

• Explicit with_pool() instead of global thread pool
• For work-stealing schedulers: with_schedule(&mut pool, DynamicScheduler)
• Access to prong metadata: for_each_with_prong(|item, prong| ...)
• NUMA-aware via distributed pools... transparent to user!

Cache-Aligned Scratch Space

Added CacheAligned<T> wrapper with 128-byte alignment to prevent false sharing in per-thread reductions:

let mut scratch: Vec<CacheAligned<usize>> =
    (0..pool.threads()).map(|_| CacheAligned(0)).collect();

(&data[..])
    .into_par_iter()
    .with_pool(&mut pool)
    .fold_with_scratch(scratch.as_mut_slice(), |acc, value, _prong| {
        acc.0 += *value;
    });
let total: usize = scratch.iter().map(|a| a.0).sum();

Performance Improvements

• Upgraded C++ optimization flags from -O2 to -O3 -march=native -mtune=native, resolving a major performance gap with Fork Union's Rust benchmarks on macOS.
• Achieved 20-30% improvement in Rust by avoiding usize pointer casts in favor of typed SyncPtr<T>, enhancing type safety and enabling more LLVM optimizations.

Updated benchmark results for N=128 bodies, 1M iterations:

| Machine | Rust FU (S) | Rust Rayon (S) | C++ FU (S) | C++ OpenMP (S) |
| -------------- | ----------: | -------------: | ---------: | -------------: |
| 16x Intel SPR | 9.8s | 🔄 32.1s | 8.7s | 12.4s |
| 12x Apple M2 | 11.0s | 🔄 1m:07.1s | 20.3s | 1m:25.9s |
| 96x Graviton 4 | 10.1s | 🔄 1m:35.6s | 26.0s | 20.8s |

Minor

• Add: CacheAligned for Rust scratch buffers (2f8b6de)
• Add: Rayon-style parallel iterators (0a66d1d)

Patch

• Fix: On 32-bit machines avoid exhausting stack in tests (de42b36)
• Make: OpenMP made optional to avoid "macro redefined" on macOS (b3df7e1)
• Make: Refresh Rayon & SPR results (0e20c4c)
• Docs: Rayon-style parallel iterators (ef0ae95)
• Make: Compile OpenMP on macOS (02983f9)
• Make: Debug 32-bit i386 with GDB (d7d89d8)
• Make: -O3 -march=native -mtune=native for C++ (6bea0e5)
• Docs: Exclude compilation from time-ing (a195fb5)
• Improve: 25% boost with typed pointers (44cd90b)