Explicitly check for None when using prefill attn_mask (#983)

When you attempt to implicitly null-check the attention_mask, you hit a
torch error:

```bash
RuntimeError: Cannot call numel() on tensor with symbolic sizes/strides
```

Simply adding an explicit check for null fixes it, and the
`--use-attention-mask` export path for prefill

sharktank/sharktank/examples/export_paged_llm_v1.py

@@ -229,7 +229,7 @@ def _(model, tokens, seq_lens, seq_block_ids, cs):
                 shard_count = llama_config.tensor_parallelism_size
 
                 tokens = ops.replicate(tokens, count=shard_count)
-                if attention_mask:
+                if attention_mask is not None:
                     attention_mask = ops.replicate(attention_mask, count=shard_count)
                 seq_block_ids = ops.replicate(seq_block_ids, count=shard_count)
                 cache_tensors = repack_cache(cs, cache_shard_dim)