GPU Pallas decode_attention improvements

[rdyro]

Changes:

• adding start_idx and kv_seq_len to support fixed-sized cache decode and sliding window decode attention
• fixing a bug where q_heads // kv_heads > block_h would cause out-of-bound indexing
• skipping work before start_idx and after kv_seq_len
• changing default sm_scale to 1 / math.sqrt(q.shape[-1]) to match jax.nn.dot_product_attention default