feat(metax): route basic Llama through canonical InfiniOps

[voltjia]

What

• Route the MetaX basic Llama/Qwen RoPE path through the canonical InfiniOps
  RotaryEmbedding API instead of the legacy InfiniOP RoPE C API.
• Enable the generated PyTorch Argmax provider in the external InfiniOps
  build on MetaX and use it for greedy sampling (top_k=1).
• Keep public InfiniCore C++ and Python interfaces unchanged.

Migration

InfiniCore route	Previous MetaX route	Canonical InfiniOps route	Alignment target
RoPE	Legacy infiniopRoPE C API	RotaryEmbedding MetaX provider	vLLM rotary_embedding
Greedy sampling (top_k=1)	Legacy random-sample path	Generated Argmax PyTorch provider	PyTorch torch.argmax

The other basic Llama routes were already migrated to canonical InfiniOps by
#1483. This PR adds the missing MetaX selection for the two routes above.

Scope

This PR targets the basic Llama architecture path used by Qwen3 with paged
FlashAttention and greedy sampling on MetaX. Non-greedy sampling, other model
families, and other attention layouts remain out of scope.

This PR is stacked on the MetaX recurrent gated delta rule build fix because
that fix is required for a complete current-main MetaX build. The Llama
runtime changes themselves are independent.

Screenshots: N/A (backend integration only).

Validation

Run on ssh metax in infiniops-ci/metax:latest on one MetaX C550, based on
main at a12eb953 plus the stacked build fix:

• Full InfiniCore MetaX release build passed with WITH_TORCH=ON and the
  generated argmax wrapper enabled.
• python3 test/infinicore/nn/rope.py --metax --debug passed 6/6 cases.
• python3 test/infinicore/ops/random_sample.py --metax --debug passed 32/32
  cases, including FP16/BF16 top_k=1 return-value and explicit-output cases.
• Qwen3-0.6B generated 64 greedy tokens and exited 0 with:

python3 examples/test_infer.py \
    --device=metax \
    --model=/workspace/Qwen3-0.6B \
    --enable-paged-attn \
    --attn=flash-attn \
    --disable-prefix-caching \
    --top-k=1 \
    --max-new-tokens=64

• An LD_PRELOAD trap covering the legacy InfiniOP execution entries used by
  the basic Transformer path did not fire during the 64-token prefill/decode
  run. The same trap reproduced the previous infiniopRoPE fallback before
  this change.
• Source scans found no InfiniLM-suffixed InfiniOps references in InfiniCore.
• The 64-token decoded output was unchanged from the pre-migration MetaX run.
• Against PyTorch Transformers greedy generation, the first 54 generated
  tokens matched exactly and token 55 diverged. This is an existing inference
  precision difference, not introduced by switching the MetaX RoPE route; the
  migration does not claim full token-for-token Transformers parity.
• clang-format 16.0.6 --dry-run --Werror and git diff --check passed.