feat(kunlun): add paged_caching, paged_attention and paged_attention_…prefill

[rubik-hua]

Enable --enable-paged-attn end to end on Kunlun XPU. Validated on P800:
all three infiniop suites pass and Qwen3-0.6B generates with the
paged-attn backend.

• paged_caching: flatten (token, head, K/V) row copies across all clusters/cores; each row moves via GM2LM/LM2GM DMAs with no cross-core sync; negative slot_mapping entries (padding) are skipped.

• paged_attention (decode): one (seq, kv_head) task per cluster; GQA heads share KV tiles staged in shared memory so paged KV is read once per kv head; online softmax keeps per-head (m, l, acc) state across tiles with fp32 accumulation; supports ALiBi and I32/I64/U32 indices.

• paged_attention_prefill (varlen, causal): one (seq, q-tile, kv_head, head-chunk) task per cluster over packed query tokens; KV tiles beyond a row's causal limit (cache_len + u) are skipped; each (head, query-token) row's online-softmax state lives in its owning core's local memory, so state updates need no shared traffic or atomics.

All three register INFINI_DEVICE_KUNLUN in the dispatch switches under
ENABLE_KUNLUN_API and validate index-dtype consistency, q/cache
contiguity, and 64-bit task/row counters. Launch defaults match the P800
topology (12 clusters x 64 cores; launching more cores than physically
present silently drops work) and are overridable via
INFINIOP_KUNLUN_{PA,PC,PAP}_{CLUSTERS,CORES}.

验证测试

python examples/test_infer.py --device kunlun --model=/mnt/geogpt-doc-new/default/infinilm-models/DeepSeek-R1-Distill-Qwen-7B --enable-paged-attn --prompt "who are you"

python examples/test_infer.py --device kunlun --model=/mnt/geogpt-doc-new/default/infinilm-models/Qwen3-0.6B --enable-paged-attn --prompt "who are you"