perf: decode common BCD fields with integer arithmetic

[maebli]

Common BCD fields now accumulate two digits at a time in u64 and convert to f64 once. Byte-aligned values through 14 digits use this path; longer values retain the original floating-point accumulation and rounding. Sign markers, byte ordering, invalid-digit errors, and positive zero are preserved.

Adds focused Criterion cases and differential coverage against the original floating-point decoder, including all 65,536 two-byte encodings and sampled values through 18 digits.

Validation:

• Application-layer tests with all features and root cargo test --all-features pass.
• Library Clippy, application-layer all-target Clippy, formatting, and the thumbv7m-none-eabi no_std build pass.
• Root all-target Clippy is blocked by existing warnings in unrelated annotation/integration tests.

Native Criterion comparison on this aarch64 Mac (opt-level=z, 30 samples, 1s warm-up, 2s measurement): 2/8/12-digit cases improved about 5%/30%/37%. Full-frame decode had no statistically detected change. The unchanged 18-digit arithmetic path measured about 5% slower, with the unrelated frame-only control also about 4% slower; native timing remains noisy. These are not MCU cycle measurements; embedded stack/flash impact remains for CI resource checks.