fix(OOM): return zeros during profiling — skip both flash_attn AND Q-tiling

flash_attn_varlen OOMs at 4096 tokens, Q-tiling also OOMs (K tensor too large).
During profiling (BI100_IN_STARTUP_PROFILE=1), return zeros immediately.
Profiling only measures memory footprint, not output correctness.

Restore: chunked_prefill=on, max_num_batched_tokens=4096.

computility-run.yaml

@@ -19,7 +19,7 @@ command:
     - --disable-log-requests
     - --disable-frontend-multiprocessing
     - --max-num-batched-tokens
-    - '256'
+    - '4096'
     - --enable-chunked-prefill
     - --max-seq-len-to-capture
     - '32768'