Add InfiniOps build integration and tensor adapters

[voltjia]

Problem

torch-infini cannot yet discover or link an installed InfiniOps build, and it
has no shared conversion layer from ATen tensors to InfiniRT TensorViews.

Scope

• Add INFINI_OPS_PREFIX plus explicit include/library directory overrides.
• Link libinfiniops before libinfinirt without adding absolute runtime
  paths.
• Add reusable dtype, device, shape, stride, and TensorView adapters.
• Add handle/config construction that accepts a native stream and explicitly
  selects native implementation index 0.
• Add build and adapter tests, but register no ATen operator in this change.

Acceptance criteria

• A focused InfiniOps build with WITH_TORCH=OFF and generated operator call
  instantiations links successfully.
• Contiguous, broadcasted, and ordinary strided tensor metadata converts
  correctly on CPU and NVIDIA.
• Unsupported dtype/device combinations fail clearly.

Depends on #2. This work can proceed in parallel with #6.