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

[voltjia]

What

• Replace the remaining covered InfiniLM-suffixed InfiniOps adapter calls with canonical InfiniOps APIs for activations, paged caching, top-k softmax, softmax, rearrange, and conv2d.
• Advance the InfiniOps submodule from 1b9c3765 to the common validated pin 21b07ebc.
• Preserve all public InfiniCore operator signatures and keep each migration as a separate commit.

Alignment

InfiniCore adapter	Canonical InfiniOps call	Open-source alignment
gelu	Gelu(input, approximate="none", out)	PyTorch torch.nn.functional.gelu, InfiniOps API
gelutanh	Gelu(input, approximate="tanh", out)	PyTorch torch.nn.functional.gelu, InfiniOps #889
relu	Relu(input, out)	PyTorch torch.nn.functional.relu, InfiniOps API
sigmoid	Sigmoid(input, out)	PyTorch torch.sigmoid, InfiniOps API
silu_and_mul	SiluAndMul(input, out)	vLLM SiluAndMul, InfiniOps API
paged_caching	ReshapeAndCacheFlash(key, value, slot_mapping, k_scale, v_scale, "auto", key_cache, value_cache)	vLLM reshape_and_cache_flash, InfiniOps #883
topksoftmax	TopkSoftmax(gating_output, bias, is_padding, renormalize, topk_weights, topk_indices, token_expert_indices)	vLLM topk_softmax, InfiniOps API
softmax	Softmax(input, dim, dtype=nullopt, out)	PyTorch torch.nn.functional.softmax, InfiniOps #890
rearrange	Copy(input, non_blocking=false, out)	PyTorch Tensor.copy_, InfiniOps #872, provider adaptation #897
conv2d	Convolution(input, weight, bias, stride, padding, dilation, false, output_padding=0, groups=1, out)	PyTorch torch.nn.functional.conv2d, ATen convolution schema, InfiniOps #882

InfiniOps uses its C++ input/attribute/output ordering. Optional top-k softmax inputs are null because the current InfiniCore API does not expose them; the required auxiliary output is plan-owned internal storage. Paged caching uses metadata-only HND-to-NHD cache views and a persistent scalar scale tensor ignored by the current auto path.

Scope

No public InfiniCore Python or C++ API changes are introduced. This consolidates the unique work from #1467, #1472, and #1476-#1479, and supersedes the earlier duplicate PRs #1468-#1471 and #1473.

The existing DeepSeek MLA paged-caching case with unequal key/value head sizes (576/512) remains unsupported by both the deprecated and canonical APIs. Extending that capability is outside this adapter-only migration.

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

Validation

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

• clang-format 16.0.6 strict check passed for all 10 changed C++ files.
• git diff --check origin/main..HEAD passed.
• Unified InfiniOps build and xmake install _infinicore passed with legacy and canonical wrappers enabled together.
• Paged caching standard equal-head-size suite: 24/24 passed across FP16/BF16/FP32 and both cache layouts.
• Top-k softmax suite: 24/24 passed.
• Sigmoid suite: 45/45 passed.
• SiLU-and-mul suite: 36/36 passed.
• Conv2d suite: 12/12 passed.