Surf dems

[qciaran]

Summary

Surface-code DEM work plus a set of loosely-grouped improvements that accumulated on the branch. The five themes:

• Surface DEM tooling: ancilla-budget-aware circuit building, batched stabilizer scheduling that mirrors Guppy's ordering, and a native-DEM threshold-sweep example (examples/surface/native_dem_threshold_sweep.py).
• Decomposed DEM with source tracking: pecos-qec::dem_builder now emits per-error-source metadata (direct vs decomposed forms, graphlike-decomposable counts) and DetectorErrorModel renders both Stim-compatible non-decomposed and decomposed outputs.
• QIS/Selene integration: implicit static qubit handles, operation-trace capture (capture_operation_trace() programmatic path plus trace_operations() JSON dump), and a physical-slot allocator in QisEngine that recycles freed slots.
• Hot-path cleanups: scratch-buffer reuse in byte_message::builder, fast single-pair path in quantum::with_flat_pairs, and reusable per-shot buffers in noise/QIS dispatch.
• Test infra: new slow marker (with a dedicated just pytest-slow lane and CI step), decomposed-DEM invariant suite, surface metadata descriptors pinned by tests, Selene parity checks across DEM/sampler/decoder paths.