Enable gpt-oss moe and mxfp4 support in OpenVINO backend

[zhaixuejun1993]

This pull request adds support for the MXFP4 quantization format to the OpenVINO backend in GGML, with special handling for 3D MoE (Mixture-of-Experts) weights and necessary updates throughout the quantization, tensor extraction, and operator support logic. Additionally, it cleans up and simplifies the handling of certain MoE-related operations for improved compatibility and numerical stability.

MXFP4 Quantization Support

• Added MXFP4 (GGML_TYPE_MXFP4) to the list of supported quantization types throughout the OpenVINO backend, including tensor creation, device support checks, and buffer allocation logic. [1] [2] [3] [4]
• Implemented MXFP4-specific extraction, packing, and conversion logic in ggml-quants.cpp, including new routines for extracting data, handling scales as E8M0, and creating the appropriate OpenVINO nodes. [1] [2] [3]
• Added special handling for 3D MXFP4 MoE weights in process_weight_tensor and related extraction/layout code, ensuring correct tensor shapes and memory layout for OpenVINO. [1] [2] [3] [4] [5]

Operator and Shape Handling Improvements

• Updated operator support checks and device compatibility logic to allow 3D MXFP4 tensors for MoE use cases, and to reject unsupported 3D quantized tensors for other types. [1] [2]
• Refined buffer allocation and tensor setting logic to recognize MXFP4 as a supported weight shape, not just 2D. [1] [2]

MoE Operation Handling Cleanup

• Removed or simplified several MoE-specific operation exclusions in is_op_unsupported_case, enabling more MoE operations to run on the OpenVINO backend and improving parity with CPU execution for numerically sensitive paths. [1] [2] [3] [4]

Codebase Maintenance

• Added necessary OpenVINO includes for new element types (float4_e2m1, float8_e8m0) to support MXFP4 quantization.

These changes collectively enable efficient and correct use of MXFP4 quantized weights, especially for advanced architectures like MoE, while also improving the backend's flexibility and maintainability.## Overview

Additional information

Requirements

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