fix(vm): compute memcmp result before translating the mutable output

[0xzrf]

Fixes #1742

Problem

sol_memcmp_ translated both input slices, then translated the mutable result pointer, and only afterwards read the inputs:

const s1 = try memory_map.translateSlice(u8, .constant, a_addr, len, check_aligned);
const s2 = try memory_map.translateSlice(u8, .constant, b_addr, len, check_aligned);
const cmp_result = try memory_map.translateType(i32, .mutable, result_address, check_aligned);
for (s1, s2) |a, b| { ... }   // reads s1/s2 after the growth

Under direct_mapping + VASA (SIMD-0460), the mutable translation can invoke the access-violation handler, which calls account.resize — reallocating and freeing the old backing

Fix

Compute the comparison into a plain i32 while both inputs are still valid, then translate the mutable result and write the value like agave

Changes

• shared/vm/syscalls/memops.zig — move the comparison above the result translation; write the precomputed value.
• Regression test memcmp reads inputs before result-triggered growth, which fails before the change (expected 0, found 153) and passes after.