[codex] Enable Kunlun Qwen2.5 inference support

[zhangyue207]

Summary

This PR contains the InfiniCore-side fixes needed to run Qwen2.5-7B-Instruct through InfiniLM on a single Kunlun XPU.

Changes

• Make Moore mate flash-attn imports optional in python/infinicore/__init__.py so non-Moore backends can import infinicore when the optional Moore Python bridge is unavailable.
• Update generic linear to materialize the transposed weight as contiguous before GEMM and apply bias after GEMM with add_.
• Add Kunlun RoPE support for I64 position IDs.

Why

During Kunlun bring-up, importing infinicore could fail because of unrelated optional Moore dependencies. Qwen2.5 inference also needed robust generic linear behavior and position-id dtype compatibility for RoPE on Kunlun.

Validation

Validated together with the matching InfiniLM changes in Docker image fixing:v1:

• Built and installed infinicore_cpp_api and _infinicore.
• Ran Qwen2.5-7B-Instruct single-card inference on Kunlun FP32.
• Verified greedy first token matches HF/CPU reference: token 114602, decoded as 昆仑.
• Verified max-new-tokens=2 output: 昆仑芯.

Runtime note: the verified path uses DISABLE_XPYTORCH=1 to avoid loading an incompatible XPU runtime from torch_xmlir.