📈 feat: add MemoryObserver for GPU-CPU communication tracking

Implement MemoryObserver to track memory transfers between GPU and CPU:
- H2D (Host to Device): CPU → GPU transfers
- D2H (Device to Host): GPU → CPU transfers
- D2D (Device to Device): GPU buffer copies
- Supports prefill/decode phase separation

Integration points in offload_engine.py:
- load_to_slot_layer: H2D with is_prefill parameter
- offload_slot_layer_to_cpu, offload_prefill_buffer_async: D2H
- write_to_prefill_buffer, write_to_decode_buffer: D2D
- load_block_sample_from_cpu, load_block_full_from_cpu: H2D

Add bench_offload.py integration for memory stats printing.

Benchmark results (Llama-3.1-8B, 64K context):
- Full Policy: Prefill H2D 262.13 GB
- XAttention: Prefill H2D 386.62 GB (1.48x)

Generated with [Claude Code](https://claude.ai/code)
via [Happy](https://happy.engineering)

Co-Authored-By: Claude <noreply@anthropic.com>
Co-Authored-By: Happy <yesreply@happy.engineering>

nanovllm/kvcache/sparse/full_policy.py

@@ -422,7 +422,7 @@ class FullAttentionPolicy(SparsePolicy):
         num_preload = min(num_slots, num_blocks)
         for i in range(num_preload):
             cpu_block_id = cpu_block_table[i]
-            offload_engine.load_to_slot_layer(load_slots[i], layer_id, cpu_block_id, chunk_idx=cpu_block_id)
+            offload_engine.load_to_slot_layer(load_slots[i], layer_id, cpu_block_id, chunk_idx=cpu_block_id, is_prefill=False)
 
         # Phase 2: Process blocks with pipeline
         for block_idx in range(num_blocks):
@@ -456,7 +456,7 @@ class FullAttentionPolicy(SparsePolicy):
             next_block_idx = block_idx + num_slots
             if next_block_idx < num_blocks:
                 next_cpu_block_id = cpu_block_table[next_block_idx]
-                offload_engine.load_to_slot_layer(current_slot, layer_id, next_cpu_block_id, chunk_idx=next_cpu_block_id)
+                offload_engine.load_to_slot_layer(current_slot, layer_id, next_cpu_block_id, chunk_idx=next_cpu_block_id, is_prefill=False)
 
             # Merge with accumulated
             with torch.cuda.stream(compute_stream):