♻️ refactor: remove cross-layer pipeline and rename compute_chunked_prefill

- Remove cross-layer pipeline from OffloadEngine (saves ~1GB GPU memory for long sequences)
  - Delete layer_k/v_buffer_a/b double buffers
  - Remove start_decode_pipeline, get_decode_layer_kv, end_decode_pipeline methods
  - Remove pipeline state tracking variables
- Simplify decode to use ring buffer pipeline only (more efficient for long sequences)
- Rename compute_chunked_attention → compute_chunked_prefill for clarity
- Add mandatory needle test requirements: --enable-offload --input-len 32768

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

nanovllm/engine/model_runner.py

@@ -644,12 +644,6 @@ class ModelRunner:
         # Get decode start position for accumulated token tracking
         decode_start_pos = self.kvcache_manager.get_decode_start_pos(seq)
 
-        # Get prefilled CPU blocks for pipeline initialization
-        cpu_block_table = self.kvcache_manager.get_prefilled_cpu_blocks(seq)
-
-        # Start cross-layer pipeline (preloads Layer 0's data)
-        offload_engine.start_decode_pipeline(cpu_block_table)
-
         # Set up context for chunked decode
         set_context(
             is_prefill=False,
@@ -666,9 +660,6 @@ class ModelRunner:
         logits = self.run_model(input_ids, positions, is_prefill=False)
         reset_context()
 
-        # End cross-layer pipeline
-        offload_engine.end_decode_pipeline()
-
         # Only offload when block is full (pos_in_block == block_size - 1)
         # This avoids unnecessary offloading on every decode step
         if pos_in_block == self.block_size - 1: