Create KV cache input tensor only if cache len > 0 for that layer (#15042)

Summary:

The MHA branch has this logic already, add it to the other branch.

Differential Revision: D84471388

Author: sxu

examples/models/llama/static_attention.py

@@ -297,6 +297,7 @@ def __init__(
                     dtype=dtype,
                 )
                 for layer_id in range(config.n_layers)
+                if cache_lens[layer_id] > 0
             }
             self.v_caches = {
                 StaticKVCache.calculate_cache_key(layer_id, 0): torch.zeros(
@@ -307,6 +308,7 @@ def __init__(
                     dtype=dtype,
                 )
                 for layer_id in range(config.n_layers)
+                if cache_lens[layer_id] > 0
             }
 
         self.config = config