Pre-fault large allocations to reduce PTE lock contention

[zkfriendly]

Pre-fault large allocations to reduce PTE lock contention

Use madvise(MADV_POPULATE_WRITE) and MADV_HUGEPAGE on large allocations (>=128 KiB) so that rayon worker threads don't all hit anonymous page faults simultaneously, avoiding contention on the kernel's PTE spinlock.

Changes

• Added prefault() in the global allocator that hints huge pages and pre-faults writable pages for large allocations on Linux.
• Uses runtime page size via sysconf(_SC_PAGESIZE) instead of a hardcoded value.
• Overflow-safe alignment arithmetic; only advises pages fully within the owned allocation.
• No-op on non-Linux targets.

Benchmark results

Benchmark 1: prefault
  Time (mean ± σ):      6.984 s ±  0.187 s    [User: 60.346 s, System: 1.147 s]
  Range (min … max):    6.730 s …  7.281 s    10 runs

Benchmark 2: base
  Time (mean ± σ):      7.902 s ±  0.096 s    [User: 63.198 s, System: 2.212 s]
  Range (min … max):    7.680 s …  8.039 s    10 runs

Summary
  prefault ran
    1.13 ± 0.03 times faster than base

~13% wall-clock improvement, ~1 s less system time (fewer kernel page faults).

perf stat comparison

Metric	prefault	base	Delta
page-faults	24,595	119,805	4.9x fewer
dTLB-load-misses	2,618,368	5,210,664	2.0x fewer
dTLB-store-misses	1,136,056	3,069,966	2.7x fewer
sys time	1.98 s	2.65 s	25% less
user time	56.44 s	58.12 s	2.9% less
wall-clock	8.57 s	9.44 s	9.2% faster

Pre-faulting eliminates ~95k minor faults that would otherwise happen in parallel across rayon threads, which is what drives the system-time reduction.
The TLB miss reduction comes from MADV_HUGEPAGE consolidating pages into 2 MiB mappings.