Feat/quantization mode

[ptoupas]

Purpose

Introduce a new quantization related feature to the modelconverter tool by adding an optional, high-level configuration parameter (quantization_mode) to simplify model quantization during RVC4 convertion.
The goal is to make quantization strategies more accessible through the UI, while still allowing full control through advanced flags.

Specification

• Introduced a new quantization_mode parameter for RVC4 model conversion, which has 5 available options: [INT8_STANDARD, INT8_ACCURACY_FOCUSED, INT8_INT16_MIXED, FP16_STANDARD, CUSTOM]. Defaults to INT8_STANDARD and the current default behavior of the quantized conversion should not be affected.
• When this parameter is set to anything but CUSTOM it will take precedence over any user-defined SNPE flags provided via the snpe_onnx_to_dlc_args, snpe_dlc_quant_args, and snpe_dlc_graph_prepare_args.
• Removed the compress_to_fp16 parameter which is now replaced by the FP16_STANDARD mode in the quantization_mode parameter.
• When the quantization_mode is set to INT8_INT16_MIXED a .json encodings file is automatically generated to keep the inputs and outputs of the model into 8-bit precision so that the generated DLC can be executed and be compatible with DAI.
• The remaining RVC4 related parameters such as use_per_channel_quantization, use_per_row_quantization, optimization_level, and others, should not be affected by the changes introduced in that PR.

Dependencies & Potential Impact

None / not applicable

Deployment Plan

None / not applicable

Testing & Validation

Tested the conversion of the following models under different configurations and settings through the CLI.

• ConvNeXt
• YOLOv6 Nano
• YuNet
• Paddle Text Recognition
• ResNet18 (local model)
• SMv4 (local model)

[klemen1999]

Left some minor comments but generally LGTM.