refactor(ops): use InfiniOps Convolution for conv2d

[voltjia]

What

• Replace ConvInfinilm with the canonical InfiniOps Convolution API.
• Supply the existing conv2d contract explicitly: transposed=false, zero output_padding, and groups=1.
• Keep input, weight, optional bias, stride, padding, dilation, and output handling unchanged.

Alignment

InfiniCore wrapper	InfiniOps call	Alignment basis
conv2d	Convolution(input, weight, bias, stride, padding, dilation, transposed=false, output_padding=0, groups=1, out)	torch.nn.functional.conv2d(input, weight, bias, stride, padding, dilation, groups), ATen convolution schema, InfiniOps Convolution, InfiniOps #882

The call follows InfiniOps' C++ input/attribute/output ordering and maps the non-transposed conv2d subset onto the full PyTorch/ATen convolution contract.

Why

The old InfiniLM convolution provider now backs the canonical Convolution operator, so the deprecated ConvInfinilm class is no longer required by InfiniCore.

Scope

This is stacked on #1478. It does not expand conv2d feature support or change the public InfiniCore API.

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.
• Existing conv2d NVIDIA suite: 12/12 passed, covering FP16/FP32, optional bias, scalar/tuple stride, padding, dilation, and strided input.
• 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Superseded by #1480, which consolidates the canonical InfiniOps adapter migrations while preserving each logical change as a separate commit.