[codex] add checkpoint replay and shared primal ownership

[shinaoka]

Summary

• add first-class checkpoint replay with CheckpointRecipe, ReplayResult, and Tape::record_checkpointed_op
• execute pullback and HVP through execution-local replay contexts, with demand-driven HVP tangents and phase-local replay
• share retained primals between graph nodes and attached TrackedValue handles so retained ops no longer require V: Clone

Root Cause

tidu previously assumed reverse execution always used permanently materialized rules. The checkpoint replay refactor fixed that for pullback/HVP, but the first cut duplicated retained primals between the tape and TrackedValue, which introduced unnecessary V: Clone requirements on retained operations.

Impact

• checkpointed nodes now replay lazily during pullback and HVP without persisting replay state on the tape
• retained nodes reuse shared primal storage while checkpointed outputs still keep their forward value available
• README, crate docs, and plan docs document the replay and shared-ownership model

Validation

• cargo fmt --all --check
• cargo clippy --workspace
• cargo nextest run --release --workspace --no-fail-fast
• cargo test --doc --release --workspace
• cargo llvm-cov nextest --workspace --release --json --output-path coverage.json
• python3 scripts/check-coverage.py coverage.json
• cargo doc --workspace --no-deps
• python3 scripts/check-docs-site.py
• bash scripts/build_docs_site.sh

Generated with Codex.