add minimax m3 attention ops on A3

[vstone-w]

Add minimax M3 adaptation operators. The list of newly added operators is as follows:

OpName	Type	Description
npu_sparse_attention_score	AscendC	decode attention,borrowed from vllm
swiglu_oai_quant	triton	Fuse SwiGLU‑OAI and int8 quantization
flash_decode_bnsd_with_topk_idx	triton	indexer used in decode
flash_prefill_bnsd_with_topk_idx	triton	indexer used in prefill
flash_decode_bnsd_with_gqa_share_sparse	triton	attention used in decode
flash_prefill_bnsd_blockq_sparse_fia	triton	attention used in prefill

These operators are integrated into the SGLang framework and are used by MiniMax M3. The overall performance test results are as follows:

Accuracy results are as follow:

[sglang-npu-bot]

/tag-and-rerun-ci