perf(demux): strided plain CRF decode; fix CNN GPU ordinal and lock policy

[jayhesselberth]

Closes #475

Shipped

• Item 2 (strided plain decode, decode_with_strided): GPU encoder + CPU decode, --no-ref-scores, 20k dual-axis, A30 node, 4 interleaved reps (rep 1 warm-up): wall 15.2-15.7 s -> 14.0-14.1 s, CPU 102.5 s -> 84.7 s, trace encode+decode 12.8 s -> 11.7 s. evaluate_demux_crf.py: 0 of 20,000 differ on both axes. Run-to-run noise ~0.2 s.
• Item 3: cuda_total_mem_bytes takes the configured ordinal. Test batch_budget_queries_the_configured_ordinal (fails pre-fix).
• Item 4: policy = recover with into_inner (as encoder_gpu.rs). Test poisoned_session_lock_is_recovered (panics pre-fix).

Dropped

• Item 1 (persistent lattice scratch): default fused GPU path, same harness, 4 reps: wall 7.7 s vs 7.6-7.9 s, trace encode+decode 5.2 s vs 5.1-5.2 s. No measurable change (20k reads is ~40 batches), and it would pin ~314 MB of device memory. Reverted.

Item 2 only applies when the GPU lattice decode is off (ESCAPEPOD_CRF_GPU_DECODE=0 or kernel load failure); the default GPU path is unchanged. Parity guard (#454) untouched.