feat(cuinterpose): multicast tracking and reconstruction

[galletas1712]

Summary

Refs #295 (approved).

Track and reconstruct multicast objects using the common resource registry and virtual-handle namespace.

Review boundary

• memory/multicast.rs owns object lifetime, device membership, bindings, unlocked collective operations, and reconstruction.
• CUDA handlers keep API-level policy. Pin/call/relock/revalidation stays inside multicast operations; handlers do not coordinate that sequence.
• This PR adds the multicast resource variant and integrates common release, mapping, import/export, inspection, and checkpoint dispatch.
• Unicast and multicast share one virtual-handle namespace and allocation-ID registry, while keeping their distinct CUDA metadata.
• Reconstruction remains ordered by creator, importer, device, and binding barriers. No new discovery or rendezvous messages.

Validation

On the assembled stack, make -C agent/cmd/cuinterpose test passes formatting, strict Clippy, GNU/musl builds, ABI/artifact checks, Rust unit tests, and packaged fake-driver tests for initialization, memory IPC, multicast, checkpoint/restore, reciprocal peer requests, and fork cleanup.

This stack retains source-only intermediate PRs; the complete Rust backend is assembled in #331, with packaging in #333 and the integration harness in #338.

Physical-GPU validation on tx5tk: 4 passed, 0 skipped on two NVIDIA B200s (driver 595.58.03). This exercised multicast/graph reconstruction, shared POSIX memory save/restore and bytes, raw-import refusal, and report parsing. Tested packaged artifacts were SHA-256 matched to the local build. The existing test image supplied Python 3.12.3, PyTorch 2.13.0+cu130, and pytest 9.1.1; this was not a fresh install of the suite's pinned Python environment. This is same-node CUDA checkpoint/restore validation, not full CRIU cross-node or framework qualification.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

---

_{Comment @coderabbitai help to get the list of available commands.}

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.