Mainline Chm02 CUDA bring-up fixes, tests, and timing surface

[tpn]

Summary

Mainline the validated Chm02 CUDA bring-up work from the gpu-batched-peeling-poc branch into main as a focused integration slice.

Scope

• Fix correctness blockers in the legacy Chm02 CUDA path.
• Keep the scope to source changes, tests, and timing/CSV surfaces.
• Exclude the standalone GPU peeling POC and internal working ledgers.

Proposed contents

• Graph.cu / GraphCu.c fixes for single-graph CUDA add-keys, peel/order capture, assignment, and verify.
• Chm02 / Chm02Compat completion-path fixes needed for known-seed CLI bring-up on Linux.
• Linux file-work context fix for the Chm02Compat path.
• Minimal GPU verify path for CUDA Chm02.
• First-class regression tests for CUDA Chm02 known-seed and generated-33000 coverage.
• Explicit CUDA timing fields in CSV output.

Why now

This slice appears mature enough to stand on its own:

• known-good Chm02 CUDA regressions pass
• Linux no-file-io and file-io bring-up issues are understood
• timing fields exist for measurement
• the work is useful independently of the newer standalone batched GPU POC

Non-goals

• mainlining the standalone experiments/gpu_batched_peeling_poc POC
• mainlining the benchmark runner and research notes
• large-scale GPU constructor/perf work

Acceptance

• existing CPU tests still pass
• CUDA Chm02 regression tests pass when CUDA is enabled
• timing fields for CuAddKeys, CuIsAcyclic, CuAssign, and CuVerify are emitted in CSV