refactor(ops): route Llama paths through InfiniOps

[voltjia]

Summary

• Enable the canonical InfiniOps flash_attn_varlen_func and flash_attn_with_kvcache providers for MetaX prefill and decode.
• Replace the legacy select_last_token_hidden provider on NVIDIA and MetaX with canonical InfiniOps add and index_select calls.
• Select the required native, generated ATen, and linked implementations through the existing InfiniOps ops.json mechanism.

Motivation

The basic Llama/Qwen inference path still bypassed canonical InfiniOps for MetaX FlashAttention and last-token selection. This change keeps those paths in InfiniCore but routes their operator work through open-source-aligned InfiniOps interfaces, without calling any Infinilm-suffixed InfiniOps operator.

Dependency

• Depends on feat(linked): add MetaX FlashAttention providers InfiniOps#934 for the MetaX linked implementations of the two canonical FlashAttention operators.
• Uses the ops.json implementation selection added by feat(build): select operator implementations from ops.json InfiniOps#931; no separate linked-provider selector is introduced.

Implementation

• multi_head_attention_varlen accepts MetaX in its existing InfiniOps eligibility check.
• mha_kvcache accepts MetaX in its existing InfiniOps eligibility check.
• select_last_token_hidden computes input_offsets[1:] - 1 with canonical add, then gathers rows with canonical index_select.
• The scalar one is initialized with a host-to-device copy during planning; no legacy InfiniOp operator is used by this route.
• The external InfiniOps build writes a build-local ops.json: FlashAttention uses linked slot 16, argmax and index_select use generated ATen slot 8, and operators that also have unrelated linked providers remain on native slot 0.

Validation

• Full infinicore_cpp_api and _infinicore builds passed on MetaX with the dependent InfiniOps branch.
• The generated linked manifest contains only the MetaX flash_attn_varlen_func and flash_attn_with_kvcache source files.
• Qwen3-0.6B paged FlashAttention inference passed after the ops.json migration with a 2-token smoke run.
• Earlier runtime tracing of the same provider code hit:
  • Operator<FlashAttnVarlenFunc, Device::Type::kMetax, 16> through the linked Torch adapter.
  • select_last_token_hidden_impl::infiniops::run.
  • Operator<FlashAttnWithKvcache, Device::Type::kMetax, 16> through the linked Torch adapter.
• A source scan over src/infinicore and include/infinicore returned no Infinilm or _infinilm references.
• A full repository scan returned no INFINI_OPS_LINKED_OPS references.
• All changed C++ files pass clang-format 16.0.6 --dry-run --Werror with the repository .clang-format.
• git diff --check origin/main...HEAD passes.

Scope

This PR only changes the basic Llama/Qwen routes described above. Other models, layouts, and fallback behavior are intentionally unchanged.