feat(api): add resumable staged Arrow and Parquet graph-import sessions

[DecisionNerd]

Problem

Large graph construction needs a bounded-memory, resumable public API. Existing batch calls repeatedly validate against global state and do not define durable staging, checkpoint, cancellation, or atomic multi-file publication semantics.

Objective

Add Rust-facade import sessions for Arrow and Parquet node/edge sources that can resume safely and publish atomically.

Requirements

• Define begin, append/register source, validate, checkpoint, resume, commit, abort, and status operations.
• Stream record batches/files with explicit memory and concurrency budgets.
• Use persistent UUID membership indexes for endpoint and duplicate validation.
• Persist a versioned journal/manifest sufficient to resume after process interruption.
• Make commit atomic: no partial graph or index publication is visible.
• Reject incompatible schemas, corrupt inputs, path traversal, symlink escapes, and resource-limit violations with typed errors.
• Make abort and stale-session cleanup deterministic and observable.

Acceptance criteria

• Multi-file node and edge imports do not materialize the complete dataset.
• An interrupted session resumes without duplicating committed work.
• Commit publishes graph data, metadata, and required indexes as one recoverable state transition.
• Abort leaves the prior project readable and removes or quarantines staged artifacts.
• Progress includes rows/bytes/files accepted, rejected, pending, elapsed, and bounded resource usage.
• Public API docs and small end-to-end tests cover Arrow and Parquet.

Scenarios

• Given a process stops after several checkpoints, when the session resumes, then it continues from durable progress without duplicate live rows.
• Given an edge references a missing node, when validation runs, then commit is refused with bounded deterministic diagnostics.
• Given cancellation during commit preparation, when the project reopens, then only the prior published graph is visible.

Non-goals

Remote object-store orchestration, domain-specific loaders, or distributed ingest.

Related issues

Canonical tracker: #735. Direct prerequisite: #737.

[DecisionNerd]

M6 #748 freezes the shared publication vocabulary that this issue should consume rather than redefine: stage, validate, durable generation, linearize, acknowledge (includes project-root directory flush), publish, abort, and recover. Normative text: ADR 0018 and tests/contracts/durability-isolation-matrix.json (graphforge-durability-isolation/1).