Add differential decode test harness

Summary

Adds a self-contained differential decode harness for catching behavioral regressions in the decoder. ZydisDiffDump decodes an input file at every byte offset across a fixed set of decoder configurations and emits per-block FNV-1a hashes of the full decode result (status, instruction struct, operands, and the pointed-to CPU/FPU flag records). Two builds of the tool over the same input are guaranteed to produce byte-identical output, so any semantic drift between a baseline and a candidate build shows up as a hash mismatch.

Intention

Decoder refactors (e.g. the performance work) need a cheap, high-coverage way to prove that a change is behavior-preserving. Exhaustive per-offset decoding over a large corpus exercises far more of the decoder than the unit tests, and reducing each decode to a hash makes a full corpus comparison a single diff.

Changes

• tools/ZydisDiffDump.c - new tool with three subcommands:
  • gen <file> <MiB> - write a deterministic xorshift64 corpus so both builds hash the exact same bytes.
  • dump <file> [start end] - emit per-block hashes; the optional range prints per-offset detail lines to bisect a mismatch down to the offending offset.
  • bench <file> - simple decode throughput benchmark (instr-only and full).
• CMakeLists.txt - builds ZydisDiffDump when ZYDIS_BUILD_TOOLS and ZYDIS_FEATURE_DECODER are enabled.
• tests/differential/diffcheck.sh - gate script that diffs the current build's dump against reference dumps captured from a baseline build (kept untracked in .diffref/) and reports the first differing lines on mismatch.

Testing

• cmake -B build -DZYDIS_BUILD_TOOLS=ON -DCMAKE_BUILD_TYPE=Release + cmake --build build --target ZydisDiffDump - builds clean.
• ZydisDiffDump gen corpus.bin 1 then ZydisDiffDump dump corpus.bin - runs across all five configs; two successive dumps of the same corpus are byte-identical (determinism confirmed via diff).

Notes

• The pointer members cpu_flags / fpu_flags reference static tables whose addresses differ between binaries, so their contents are hashed and the pointers nulled before hashing the instruction struct - otherwise identical decodes would hash differently across builds.
• Reference dumps live untracked in .diffref/; the harness compares builds, it does not ship a golden baseline.
• Opened as a draft for review of the approach.