refactor(ops): use canonical InfiniOps activation APIs

[voltjia]

What

• Advance InfiniOps from 1b9c3765 to 21b07ebc.
• Replace the InfiniLM-suffixed GELU, GELU-tanh, ReLU, sigmoid, and SiLU-and-mul adapter calls with their canonical InfiniOps APIs.
• Preserve every InfiniCore public signature and execution path.

Alignment

InfiniCore wrapper	InfiniOps call	Alignment basis
gelu	Gelu(input, none, out)	torch.nn.functional.gelu(input, approximate=none), InfiniOps Gelu
gelutanh	Gelu(input, tanh, out)	torch.nn.functional.gelu(..., approximate=tanh), InfiniOps #889
relu	Relu(input, out)	torch.nn.functional.relu(input, inplace=False), InfiniOps Relu
sigmoid	Sigmoid(input, out)	torch.sigmoid(input), InfiniOps Sigmoid
silu_and_mul	SiluAndMul(input, out)	vLLM SiluAndMul, InfiniOps SiluAndMul

InfiniOps follows its C++ convention of input tensors, attributes, then output tensors; this PR only changes the adapter target.

Why

The canonical providers are now available in InfiniOps, so these adapters no longer need the deprecated *Infinilm compatibility classes.

Scope

No Python/C++ public InfiniCore API, tensor layout, or kernel behavior changes. The InfiniOps pin also contains the canonical APIs used by the following stacked migration PRs.

Screenshots: N/A (backend adapter migration only).

Validation

Run on ssh nvidia in accelerator-dev/nvidia:latest on NVIDIA A100 GPUs:

• Full infinicore_cpp_api and _infinicore build/install passed.
• Python extension import and dynamic linking passed with InfiniOps' private InfiniRT SONAME path.
• Existing sigmoid NVIDIA suite: 45/45 passed.
• Existing SiLU-and-mul NVIDIA suite: 36/36 passed.
• python3 scripts/format.py --ref 850fb3f7 --path src --check with clang-format 16.0.6.
• git diff --check.

[voltjia]

Superseded by #1480, which consolidates the canonical InfiniOps adapter migrations while preserving each logical change as a separate commit.