refactor(ops): migrate GELU and sigmoid to canonical InfiniOps APIs

[voltjia]

Summary

• migrate the InfiniCore GELU adapters from deprecated GeluInfinilm and GelutanhInfinilm to canonical Gelu
• select the PyTorch-compatible none and tanh approximation modes explicitly
• migrate the sigmoid adapter from deprecated SigmoidInfinilm to canonical Sigmoid
• keep the InfiniCore public APIs and graph behavior unchanged

API alignment

InfiniCore adapter	Previous InfiniOps API	Canonical InfiniOps API	Alignment target and evidence
GELU	GeluInfinilm::Call(handle, config, input, approximate, out)	Gelu::Call(handle, config, input, approximate, out)	PyTorch Python API: torch.nn.functional.gelu(input, approximate='none'). The ordinary adapter passes none; the tanh adapter passes tanh. InfiniOps adds an explicit trailing output tensor.
Sigmoid	SigmoidInfinilm::Call(handle, config, input, out)	Sigmoid::Call(handle, config, input, out)	PyTorch Python API: torch.sigmoid(input, *, out=None). InfiniOps represents the optional Python output as its required trailing C++ output tensor.

Dependency

This PR is based on refactor/migrate-paged-caching-infiniops / #1467 and advances the InfiniOps gitlink from e733e325 to 498c9aaa, the head of InfiniOps #889, which adds canonical CUDA-compatible GELU and sigmoid providers.

Against #1467, this PR changes three adapter files and the InfiniOps gitlink.

Validation

Validated commit 37b8fb8b remotely on ssh nvidia with accelerator-dev/nvidia:latest, using the exact InfiniOps #889 head 498c9aaa in a temporary integration merge:

• built and installed _infinicore with the canonical gelu and sigmoid providers in the wrapper allowlist
• ran python3 test/infinicore/ops/sigmoid.py --nvidia: 45/45 passed
• compiled and ran a temporary C++ NVIDIA smoke test through infinicore::op::gelu and infinicore::op::gelu_tanh; both outputs matched host references
• clang-format 16.0.6 --dry-run --Werror passed for all three changed C++ 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Superseded by #1480, which consolidates the canonical InfiniOps adapter migrations while preserving each logical change as a separate commit.