Avoid per-node numeric type tuples in compact decoding

[bradhilton]

Compact decoding rebuilds a (bool, int, float) tuple at every JSON node. Replace it with direct int and float checks; bool remains covered by int. Container reconstruction, string references, validation errors and independent mutable results are unchanged.

Validation: 50 focused tests pass (23 new scalar/ownership/error cases plus the existing compact serialization suite), Ruff/format and scoped type checks pass, and uv lock --check --offline passes. An independently reviewed public fixture also passed 22 control groups, including the actual compressed Caladan dataset stream's order, slices, repeated/interleaved iteration and identity.

Public CPU benchmarks on CPython 3.12.13 used 13 randomized paired rounds. With cyclic GC outside the timed sections, compact validation was 14.3% faster and an eight-row compressed stream was 12.3% faster. The initial natural-GC measurements were noisy and showed only about 3% median gains. These are local synthetic measurements, not production latency or GPU-utilization claims. The fixtures use public ChatCompletion objects; no private inputs or GPU workload.

This draft is separate from the tokenizer and checkpoint-copy changes. It adds no cache, decoder API, offload or iteration change.