NEFF compile-cache reuse in examples and benchmarks

[scttfrdmn]

Phase 3 gap per suite ROADMAP.md.

NKI kernels incur NEFF compilation cost on first call. Examples (examples/scf_eigen.py) and benchmarks (benchmarks/bench_*.py) should verify the Neuron NEFF cache is warm across runs so users don't pay for recompiles.

Scope

• Environment: NEURON_COMPILE_CACHE_URL / NEURON_CC_FLAGS --cache_dir consistently set
• scripts/run_neuron_tests.sh — export cache dir; persist across runs on the CI instance
• Examples and benchmarks log "NEFF cache HIT" vs "NEFF cache MISS" on each kernel call (optional: via env var TRNSOLVER_LOG_NEFF=1)
• Benchmarks report two numbers: cold-cache and warm-cache

Done means

• Second run of bench_solver.py with @pytest.mark.neuron is ≥ 10x faster than first run (compilation amortized)
• CI run-summary shows cache hit rate

Depends on

• Phase 3 — preconditioned iterative solvers + NEFF cache reuse #28 (Phase 3 tracker)