Make allocation and copy stream-aware

[voltjia]

Problem

The allocator and copy_ implementation are synchronous, ignore
non_blocking, and do not associate memory lifetime with the active stream.

Scope

• Use InfiniRT async allocation/free APIs where supported.
• Route copies through the current native stream.
• Honor non_blocking only when host memory and backend capabilities make it
  valid; otherwise preserve synchronous semantics or fail clearly.
• Add pinned-host-memory support needed for asynchronous host/device copies.
• Do not add computational operators in this change.

Acceptance criteria

• Nondefault-stream copies have correct ordering and lifetime.
• non_blocking=True is behaviorally tested with pinned host memory.
• Unsupported backend paths do not silently claim asynchronous behavior.

Depends on #6 and #7.