add dense_lightning_indexer_grad_klloss and other custom ops

添加npu_dense_lightning_indexer_grad_klloss，npu_dense_lightning_indexer_softmax_lse, npu_sparse_lightning_indexer_grad_klloss 三个自定义接入算子。支持pynative模式。
接口对标pta。

def npu_dense_lightning_indexer_grad_kl_loss(
        query,
        key,
        query_index,
        key_index,
        weights,
        softmax_max,
        softmax_sum,
        softmax_max_index,
        softmax_sum_index,
        scale_value,
        query_rope=None,
        key_rope=None,
        actual_seq_qlen=None,
        actual_seq_klen=None,
        layout=None,
        sparse_mode=None,
        pre_tokens=None,
        next_tokens=None)
      
    def npu_dense_lightning_indexer_softmax_lse(
        query_index,
        key_index,
        weight,
        actual_seq_qlen=None,
        actual_seq_klen=None,
        layout=None,
        sparse_mode=None,
        pre_tokens=None,
        next_tokens=None)
        
    def npu_sparse_lightning_indexer_grad_kl_loss(
        query,
        key,
        query_index,
        key_index,
        weights,
        sparse_indices,
        softmax_max,
        softmax_sum,
        scale_value,
        query_rope=None,
        key_rope=None,
        actual_seq_qlen=None,
        actual_seq_klen=None,
        layout=None,
        sparse_mode=None,
        pre_tokens=None,
        next_tokens=None)
    ```

<!-- bot3-issue-meta -->
schema_version: 1
source: gitcode
gitcode_repo: mindspore/hyper-parallel
gitcode_issue: 97
source_url: https://gitcode.com/mindspore/hyper-parallel/issues/97
<!-- /bot3-issue-meta -->

Start by locating the existing PTA counterparts and the project's custom-operator registration or entry points. Compare their interfaces with the three signatures in the issue, then determine how PyNative support is handled. Done means all three NPU operators are registered with PTA-compatible behavior and PyNative support.