perf: decode rewrite, encode zeros optimization, iai-callgrind benchmarks (0.3.1)

Decode rewritten with shrinking-slice pattern (split_first/split_at)
instead of index arithmetic. Decode zeros 14% faster, decode nonzero
flips from 5% behind corncobs to 2% ahead, decode mixed closes to tied.

Encode zeros fast path moved before split_at — now tied with cobs
(was 6% behind). Code size reduced to 821B (was 869B).

Replaced criterion with iai-callgrind for deterministic instruction-count
benchmarks via Valgrind. Zero run-to-run variance. Expanded to 162
benchmarks across 9 sizes.

ucobs now wins or ties 16 of 18 benchmarks vs corncobs and cobs.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

.gitignore

@@ -2,6 +2,8 @@
 /fuzz/target
 /fuzz/corpus
 /fuzz/artifacts
+/mutants.out
+/mutants.out.old
 Cargo.lock
 __pycache__/
 *.pyc