feat(cuda): multi-GPU runtime facade + 3-phase migration orchestrator (v1.1)

Per spec sections 3.1 and 4.1. Host-side orchestration only; GPU-side
kernels committed separately.

crates/ringkernel-cuda/src/multi_gpu/:

runtime.rs (1229 lines):
- MultiGpuRuntime: device pool + NvlinkTopology + MultiGpuRegistry
  + MigrationController
- GpuBackend trait abstracts CudaRuntime (mockable for tests)
- PlacementHint: Auto, Pinned, WithActor, NvlinkPreferred
- Peer access bookkeeping (defers cuCtxEnablePeerAccess to hardware phase)
- Cross-GPU send routing (uses topology to pick NVLink path)

migration.rs (405 lines):
- MigrationPlan, MigrationReport, PhaseDurations
- RebalanceStrategy: LoadBalance, CommunicationAware, Explicit
- MultiGpuError with RingKernelError conversion
- 3-phase orchestration: quiesce → transfer → swap

migration_controller.rs (353 lines):
- Atomic concurrency cap (default 1000)
- Global buffer budget (default 8GB) with AtomicU64 accounting
- TokenBucket rate limiter
- MigrationPermit RAII release pattern

staging.rs (383 lines):
- StagingBuffer: framed (u32 length + payload) push/drain
- Optional disk-spill overflow for bursts
- CRC-32/ISO-HDLC checksum (pre- and post-transfer)
- High-water tracking

registry.rs (220 lines):
- Bidirectional KernelId↔GPU indexing
- Per-GPU load accounting
- Atomic location updates for migration

Tests: 62 new tests + 16 topology + 6 migration_kernels = 84 multi_gpu
tests total (188 ringkernel-cuda --lib tests pass w/ 24 ignored for
hardware). All 3-phase ordering, rate limiting, budget enforcement,
checksum match/mismatch, disk spill, placement hint resolution,
rebalance source/target selection verified without hardware.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: mivertowski

crates/ringkernel-cuda/Cargo.toml

@@ -50,6 +50,7 @@ nvml-wrapper = { version = "0.12", optional = true }
 
 [dev-dependencies]
 tokio = { workspace = true, features = ["test-util", "macros", "rt-multi-thread"] }
+tempfile = { workspace = true }
 
 # Note: GPU execution tests require the `cuda` feature to be enabled.
 # Run with: cargo test -p ringkernel-cuda --features cuda