Use active sequences for container decoding

[ruslandoga]

Keeps active container state in decoder arguments and stores frames only at nesting boundaries, avoiding expanded type lists and per-item continuation allocation. Small compile-time transition macros remove repeated state plumbing without adding runtime state objects; arrays reuse one schema, tuples traverse original schema tails, and maps build directly as each pair completes. Benchmarks show arrays of 3–100 items about 1.2–2.9x faster, maps with 1–10 pairs about 2.5–5.4x faster, and three-field tuples using 12% fewer reductions and 10% less memory than master. Tested with MIX_ENV=test mix test (475 passed), and macro-versus-explicit benchmarks use identical memory across array sizes.