cuda.core: make buffer cleanup context-safe - #2455

Summary

Make device-pointer cleanup retain and activate the CUDA context required by the allocation or its deallocation stream. This prevents deferred graph attachment cleanup and ordinary Buffer destruction from leaking memory when no context is current on the cleanup thread.

Changes

• Resolve the release context in Cython (Stream_resolve_context) and pass an explicit stream/context pair into device-pointer resource handles; C++ stores that pair and makes the context current during destruction.
• Retain allocation and deallocation-stream context handles, including owning primary-context retains when the driver permits them.
• Temporarily activate the retained context for native frees, graphics unmap, and MemoryResource.deallocate() callbacks, then restore the caller's prior context.
• Stop treating CUDA_ERROR_INVALID_CONTEXT as successful memory-pool deallocation, and report cleanup failures on stderr instead of discarding them.
• Add an optional keyword-only stream= argument to Buffer.from_handle and ManagedBuffer.from_handle, defaulting to the calling thread's default stream when a memory resource owns the pointer.
• For IPC imports that arrive with a default-stream token and no current context, bind the pool device's primary context so release still has a context to run under.

Review guide

1. Review _stream.pyx / _buffer.pyx / _ipc.pyx for how Cython chooses the release context.
2. Review resource_handles.cpp/.hpp for storing that context as a handle dependency and restoring it during teardown.
3. Review the memory and graph lifetime tests for cleanup with no current context.

Related Work

• Follow-up to cuda.core: add graph definition node updates #2395.
• Related to cuda.core: reject stream=None — require explicit stream everywhere #2001 and Inconsistent lifetime management across from_handle / from_* APIs #1989.