Add x86 AVX2 and AVX512 memory-bandwidth backends

[zqb-all]

What

Adds modern x86 SIMD backends (AVX2 and AVX512) alongside the existing SSE2 code. On x86-64 the benchmark now reports SSE2, AVX2 and AVX512 side by side in a single run, gated at runtime on the CPU's actual capabilities (read from /proc/cpuinfo flags via the existing check_cpu_feature()), so older CPUs keep running only the SSE2 list and never execute an unsupported instruction.

Provenance / credits

This is not new code — it gathers existing community work and wires it into this fork:

• Joel Luth (letrout / jluth) — author of the AVX2/AVX512 assembly + benchmarks, originally proposed upstream as ssvb/tinymembench#23, based on Siarhei Siamashka's (ssvb) original SSE2 code.
• Tony Mason (fsgeek) — gcc AVX512 feature-detection help.
• Travis Downs (travisdowns/tinymembench) — hosted the x86 AVX branch this integration draws from.

Original copyright headers are preserved in x86-avx2.{S,h} / x86-avx512.{S,h}, a Credits section is added to the README, and the commit carries Co-authored-by: Joel Luth.

Design

• New files: x86-avx2.{S,h}, x86-avx512.{S,h}.
• asm-opt.c dynamically concatenates SSE2 → AVX2 → AVX512 lists based on detected features; all AVX code is guarded on __amd64__ and compiles to empty objects on other targets, so non-x86 builds are unaffected.
• The AVX headers' size_t size params were aligned to int to match the bench_info function-pointer signature used by SSE2.

Verification (Intel Xeon Gold 6338N, Ice Lake-SP)

• gcc -c compiles both backends; a standalone harness confirms every copy is bit-exact and no SIGILL.
• The real tinymembench binary lists all three tiers, runs to completion (exit 0), full latency section included.
• As expected, at 32 MB (DRAM) regular copy is flat across all three (single-core memory wall), while the nontemporal-store path scales cleanly: SSE2 6.2 → AVX2 8.8 → AVX512 11.9 GB/s, and in-L1 copy jumps SSE2 85 → AVX ~150 GB/s.

Note on scope

A survey of the fork ecosystem found the remaining community patches to be either already present in this fork's nuumio base (ARM/portability fixes — confirmed by empty cherry-picks) or not mergeable as-is (e.g. an arter97 hugetlbfs change that exit(1)s without preconfigured huge pages). AVX2/AVX512 is the one substantial, clean, non-redundant increment — hence this focused PR.