[ROCm] Fix cu_seqlens_q off-by-one in AITER FA speculative decode path (#39120)

Signed-off-by: Bortlesboat <bortstheboat@gmail.com>

vllm/v1/attention/backends/rocm_aiter_fa.py

@@ -1181,7 +1181,7 @@ class AiterFlashAttentionImpl(AttentionImpl):
                     )
 
                     descale_shape = (
-                        attn_metadata.query_start_loc[:num_decodes].shape[0] - 1,
+                        num_decodes,
                         key_cache.shape[2],
                     )
                     unified_attention(
@@ -1189,7 +1189,7 @@ class AiterFlashAttentionImpl(AttentionImpl):
                         k=key_cache,
                         v=value_cache,
                         out=output[:num_decode_tokens],
-                        cu_seqlens_q=attn_metadata.query_start_loc[:num_decodes],
+                        cu_seqlens_q=attn_metadata.query_start_loc[: num_decodes + 1],
                         max_seqlen_q=decode_max_query_len,
                         seqused_k=attn_metadata.seq_lens[:num_decodes],
                         max_seqlen_k=attn_metadata.max_seq_len,