Preserve current device across runtime operations

[voltjia]

Problem

Operations that target an explicit device currently call SetDevice and leave
that device current after returning. In particular, synchronize(device) and
tensor copies can unexpectedly mutate caller-visible device state.

Scope

• Preserve and restore the current device around explicit-device runtime calls.
• Cover success and error paths with an exception-safe guard.
• Apply the same rule to allocation, copy, and synchronization entry points.
• Do not add stream support or new operators in this change.

Acceptance criteria

• On a multi-device backend, explicit operations on device 1 leave the original
  current device unchanged.
• Same-device and cross-thread behavior remains correct.
• CPU and NVIDIA smoke tests pass.

Depends on #2.