fix: preserve local in-process channels across peer DSM detachment

[stuhood]

Problem

In multi-stage distributed queries where tasks are multiplexed across fewer worker processes than logical tasks (e.g. 8 tasks on 7 workers), peer workers can finish quickly, exit, and drop their DSM senders. This causes the remaining worker's DSM receiver to observe detachment (RecvBatchOutcome::Detached).

Previously, DrainHandle::detach_data_scope failed all registered channel buffers upon DSM detachment. When a worker hosted both a local producer and consumer fragment communicating via open_local_partition_sink, this prematurely failed the local in-process channel with:

Execution error: transport receiver detached before this channel's EOF; the producer went away

Solution

• Track this_proc on DrainHandle to distinguish local in-process channels (sender_proc == this_proc) from remote DSM channels.
• In detach_data_scope and register_data_channel, avoid failing local channels on DSM inbox detachment; local channels remain live until the local producer finishes.
• Add failed: Option<String> to ChannelBufferRegistry so that fatal errors from fail_scope continue to fail both local and remote channels.
• Implement Drop for LocalDrainPartitionSink to fail pending buffers if a local producer drops without calling finish() (e.g. due to panic or task error).
• Wire local stream cancellation through MppMesh::cancel_stream and DrainHandle::cancel_stream.