vfio-user: implement migration v2 (device-feature + DMA logging)

[maxpain]

Summary

This adds the vfio-user migration v2 wire protocol to the vfio-user crate, byte-compatible
with libvfio-user's include/vfio-user.h and the kernel VFIO_DEVICE_FEATURE migration uAPI. It is
the foundational piece for VMM-driven live migration of vfio-user devices — the same
VfioMigrationState semantics the kernel-VFIO path already uses, carried over the socket.

What's added

• New commands: VFIO_USER_DEVICE_FEATURE (16), VFIO_USER_MIG_DATA_READ (17),
  VFIO_USER_MIG_DATA_WRITE (18), plus the VFIO_DEVICE_STATE_* / VFIO_DEVICE_FEATURE_* /
  VFIO_MIGRATION_* constants.
• Server dispatch for the device-state FSM, MIG_DATA streaming (a short read marks
  end-of-stream, per the protocol), and DMA dirty logging. Wire-controlled allocations are bounded.
• ServerBackend gains default-erroring hooks (migration_flags,
  migration_{set,get}_state, migration_{read,write}_data, dma_logging_{start,stop,report}).
  The backend owns FSM transition validity — the server forwards the requested target state to it,
  the same way vfio-ioctls does for kernel VFIO.
• Client gains the matching methods so a VMM can drive migration.
• An integration test driving the full FSM + MIG_DATA round-trip + DMA-logging bitmap over a
  socketpair.

Non-migratable backends are unaffected: the new commands stay rejected.

Testing

• cargo test, cargo fmt -- --check, and cargo clippy --all-features are clean.
• The new integration test raises workspace line coverage (~60% → ~65%), so
  coverage_config_x86_64.json is updated accordingly.

Context

This lets a VMM live-migrate a vfio-user device (e.g. an emulated NVMe controller) using the
migration-v2 model that landed for kernel VFIO. A Cloud-Hypervisor consumer
(Migratable for VfioUserPciDevice) is a follow-up that builds on this.