OpenVINO Backend: Support common MTMD ops, fix input tensor creation bugs

[mostafafaheem]

Overview

• Supports new ops UPSCALE, SWIGLU_CLAMP, SIN, COS, LOG, MEAN, SUM, GELU, GELU_ERF, GELU_QUICK, ELU, HARDSWISH, HARDSIGMOID, STEP, ABS, SGN, FLOOR, CEIL, ROUND, EXPM1 - commonly used in multimodal LLMs
• Supports ROPE types MROPE and VISION
• Ensures input types are equal in translate_1to1_match_2_inputs
• Moves all UNARY ops into a single shared file
• Fixes bug in ggml_backend_openvino_buffer_init_tensor where tensor views reused the parent's ov::Tensor, causing incorrect input tensor shapes (observed in llama-mtmd-debug).
• Adds more sanity checks in convert_ggml_input_to_ov, enables contiguous repack for non-split models.

Additional information

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: