[Bugfix] Fix stale common_attn_metadata.max_seq_len in speculative decoding with Eagle (#32312)

Signed-off-by: Ofir Zafrir <ofir.zafrir@intel.com>

vllm/v1/spec_decode/eagle.py

@@ -466,6 +466,12 @@ class EagleProposer:
             # For the requests that exceed the max model length, we set the
             # sequence length to 1 to minimize their overheads in attention.
             common_attn_metadata.seq_lens.masked_fill_(exceeds_max_model_len, 1)
+            # Increment the maximum sequence length. We increment max_seq_len
+            # unconditionally even though some seq_lens may have been capped above,
+            # as max_seq_len serves as an upper bound for sequence lengths.
+            common_attn_metadata.max_seq_len = min(
+                common_attn_metadata.max_seq_len + 1, self.max_model_len
+            )
 
             # Also update the CPU-side shadow; NOTE: this is hacky and should be
             # removed in when common_attn_metadata.seq_lens_cpu is deprecated.