libfabric: Use desc-specific target offset (ai-dynamo#883)

This fixes a bug in multi-descriptor transfers where descriptors
point to different offsets within the same registered memory region.

Without this fix, RDMA reads always target offset 0. Should extract
each descriptor's specific target address instead.

Also impacted: Block-based transfers (Iteration N would read blocks
from iteration 0, etc), Partial buffer updates, etc.

Signed-off-by: Tushar Gohad <tushar.gohad@intel.com>
(cherry picked from commit f62d1d7)

Author: tsg-

src/plugins/libfabric/libfabric_backend.cpp

@@ -1035,7 +1035,8 @@ nixlLibfabricEngine::postXfer(const nixl_xfer_op_t &operation,
         int gpu_id = local[desc_idx].devId;
 
         NIXL_DEBUG << "Processing descriptor " << desc_idx << " GPU " << gpu_id
-                   << " addr: " << transfer_addr << " size: " << transfer_size;
+                   << " local_addr: " << transfer_addr << " size: " << transfer_size
+                   << " remote_addr: " << (void *)remote[desc_idx].addr;
 
         NIXL_DEBUG << "DEBUG: remote_agent='" << remote_agent << "' localAgent='" << localAgent
                    << "'";
@@ -1071,11 +1072,14 @@ nixlLibfabricEngine::postXfer(const nixl_xfer_op_t &operation,
         }
 
         // Prepare and submit transfer for remote agents
+        // Use descriptor's specific target address
+        uint64_t remote_target_addr = remote[desc_idx].addr;
+
         nixl_status_t status = rail_manager.prepareAndSubmitTransfer(
             op_type,
             transfer_addr,
             transfer_size,
-            remote_md->remote_buf_addr_,
+            remote_target_addr,
             local_md->selected_rails_,
             local_md->rail_mr_list_,
             remote_md->rail_remote_key_list_,