Fix silent feature corruption: register-as-number parsing, x64 negative masks, duplicate string/stack-string features

[mnaza]

Closes #24.

Four extraction bugs that silently emit wrong or duplicate features. 3 regression tests added (16 pass, was 13).

1. Registers ah/bh/ch/dh and hex-looking labels parsed as numbers

parse_operand_to_number case 2 stripped the h suffix and parsed the rest as hex: "ah" → "a" → Some(10), so mov al, ah emitted Number(0xA). Case 5 (bare hex fallback) also accepted labels like beef/face. Both paths now require a leading digit — the Intel convention for hex literals (0ABh keeps working; ah, beef do not parse).

2. Negative immediates masked to 32 bits on x64

(s as u32) as i128 — mov rax, -1 emitted Number(0xFFFFFFFF) instead of Number(0xFFFFFFFFFFFFFFFF). New mask_to_bitness helper follows the function's bitness.

3. stack string pushed once per remaining instruction

The push sat inside the instruction loop with no break, so a block crossing the threshold got one duplicate characteristic per remaining instruction. Now emitted once per basic block.

4. Every ASCII string emitted twice

extract_file_strings calls extract_ascii_strings and extract_unicode_strings — but the latter also ran a plain-ASCII pass ([\x20-\x7E]{4,}, the same printable class). Each ASCII string landed twice in the string feature set: double matching work for every string rule and duplicate VAs in feature maps. The Unicode extractor is now UTF-16-only; ASCII coverage stays with extract_ascii_strings (whose class is slightly wider — includes \t).

Out of scope (noticed while testing, pre-existing)

Case 3 of parse_operand_to_number (-0x10 / +0x10 sign forms) is dead code: RE_NUMBER_HEX captures the number with the 0x prefix, and i128::from_str_radix rejects it — so signed hex immediates never parse. Left alone here to keep the PR focused; happy to file a follow-up if you want it fixed.

Clippy clean, cargo fmt applied, CHANGELOG entry under [Unreleased]. Based on master; independent of the other open PRs (#17 touches the same file but different regions; only the [Unreleased] CHANGELOG heading is shared).

[marirs]

All 4 pull requests has conflicts

[mnaza]

Corrected, now all 4 mergable

[marirs]

conflicts..
if this is the latest - lets not fix the othe PR's

[marirs]

still

[mnaza]

Now should merge