fix(kunlun): memset fallback, graph-op destructor UB, missing selectlast_token_hidden kernel

[rubik-hua]

Kunlun inference was broken by three independent issues; this commit
fixes all of them.

1. Kunlun memset support (src/infinirt/kunlun/infinirt_kunlun.cc)

memsetDevice/memsetDeviceAsync returned INFINI_STATUS_NOT_IMPLEMENTED.
This surfaced after fafd723 (issue/1170), which made Tensor::zeros
actually zero the buffer via memset; previously the buffer from empty()
was left untouched, so the stub was never exercised. XRE exposes no
memset API, so emulate it with chunked host-to-device copies from a
16 MiB host staging buffer, keeping host memory bounded regardless of
fill size. The async variant falls back to the sync copy, consistent
with mallocAsync/freeAsync already falling back to sync XRE calls.

2. Uninitialized DispatchableGraphOperator members (include/infinicore/graph/graph.hpp)

planned_meta_/runner_/deleter_ had no default initializers. When a
derived constructor throws (e.g. an operator unsupported on the current
device), stack unwinding still runs the base destructor, which read the
garbage deleter_ and jumped through it - a silent SIGSEGV that masked
the real "operator not supported" error and made the process exit with
no traceback. Default-initialize all three to nullptr so the existing
null check in the destructor skips cleanup safely. No functional change
for already-supported devices: every constructor overwrites the three
members via INFINICORE_GRAPH_OP_DISPATCH before they are read.

3. Kunlun kernel for select_last_token_hidden for already-supported devices: every constructor overwrites the three members via INFINICORE_GRAPH_OP_DISPATCH before they are read.

4. Kunlun kernel for select_last_token_hidden

Add an XPU kernel (f16/bf16/f32) and register it in the infiniop
dispatcher, unblocking the static-cache path on Kunlun. The build picks
the new files up through the existing ops//kunlun/.xpu glob; no build
change needed. Note that paged_caching/paged_attention still have no
Kunlun kernels, so --enable-paged-attn remains unsupported on this
platform for now.

Also fix an out-of-scope index in a cat.cc assert (tensors[i] -> tensor,
where no such i exists in that loop), which broke debug builds.

当前最新代码仓上昆仑芯推理报错：

还会有跟graph相关的coredump，堆栈如下：

修复后验证：
python examples/test_infer.py --device kunlun --model=/mnt/geogpt-doc-new/default/infinilm-models/Qwen3-0.6B --tp 1 --batch-size 1 --prompt "who are you"

[wooway777]

agent的结论，我没机器验证不了

[wooway777]

另外static cache的推理流程中有cpu交互，预期就是不能跑graph。需要补充paged attention或者flash attention相关的支持。

[rubik-hua]

另外static cache的推理流程中有cpu交互，预期就是不能跑graph。需要补充paged attention或者flash attention相关的支持。

paged attention或者flash attention 我在实现，后面我单独提PR。 不过没编译graph，运行也没指定 --enable-graph情况下，会coredump在graph代码中，这个肯定是bug，需要修复的。 上面agent给的检视意见我会过一遍修复掉。

[rubik-hua]

agent的结论，我没机器验证不了

这三个问题已经修改

[wooway777]