refactor(ops): migrate activation adapters to canonical InfiniOps APIs

[voltjia]

Summary

• migrate the InfiniCore ReLU adapter from deprecated ReluInfinilm to canonical Relu
• migrate the InfiniCore SiLU-and-multiply adapter from deprecated SiluAndMulInfinilm to canonical SiluAndMul
• keep the InfiniCore public API, dispatch behavior, and pinned InfiniOps submodule unchanged

API alignment

InfiniCore adapter	Previous InfiniOps API	Canonical InfiniOps API	Alignment target and evidence
ReLU	ReluInfinilm::Call(handle, config, input, out)	Relu::Call(handle, config, input, out)	PyTorch Python API: torch.nn.functional.relu(input, inplace=False). InfiniOps retains its C++ input/output convention while using the canonical operator name and unary input contract.
SiLU and multiply	SiluAndMulInfinilm::Call(handle, config, input, out)	SiluAndMul::Call(handle, config, input, out)	vLLM Python API: SiluAndMul.forward(x) -> Tensor, with x[..., :d] gated by SiLU and multiplied by x[..., d:]. InfiniOps exposes the same single-input contract with an explicit output tensor.

The currently pinned InfiniOps commit (1b9c37659211eeaeedb48b9691f79e4c9e4412f9) already provides both canonical operators, so this PR does not update the submodule.

Validation

Validated commit 0dbf6e0d remotely on ssh nvidia with accelerator-dev/nvidia:latest:

• built and installed _infinicore from a clean checkout with the exact pinned submodules
• confirmed libinfinicore_cpp_api.so resolves libinfiniops.so from the exact installed package
• ran python3 test/infinicore/ops/silu_and_mul.py --nvidia: 36/36 passed across FP16, BF16, FP32, functional, and out-parameter cases
• compiled and ran a temporary C++ smoke test through infinicore::op::relu on NVIDIA; device execution and host result comparison passed
• clang-format-16 --dry-run --Werror passed for both changed files
• git diff --check passed

[voltjia]

Superseded by #1480, which consolidates the canonical InfiniOps adapter migrations while preserving each logical change as a separate commit.