refactor: remove EncodeParallel after benchmarks show 14% slower

Parallel encode was 14% slower on Bitcoin, 5-12x slower on small inputs,
and allocated 7x more memory. Per-match dictionary lookups are too fast
for parallelization overhead. Remove from public API and benchmarks.
Update benchmark results (93 benchmarks, clean run without parallel).

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: HavenDV