Fix mixed precision calibration encodings

[ptoupas]

Purpose

Fix an issue with the custom encodings generation for mixed-precision (W8A16) quantization on RVC4 and introduce a new quantization mode, INT8_INT16_MIXED_ACCURACY_FOCUSED, which combines mixed-precision quantization with the enhanced calibration method to prioritize accuracy.

Specification

• Added a new function to retrieve tensors preceding the model outputs up to a specified depth and include them in the custom encodings generated for mixed-precision (W8A16) quantization on RVC4. This serves as a temporary workaround to maintain compatibility with DAI, which does not support INT16 inputs and outputs.
• Introduced a new quantization mode, INT8_INT16_MIXED_ACCURACY_FOCUSED, which combines mixed-precision quantization with enhanced calibration to improve accuracy.

Dependencies & Potential Impact

None / not applicable

Deployment Plan

None / not applicable

Testing & Validation

Tested on a pool of different models, like yolov8n-seg, yolov10n, yolov11n, and tasks like classification, instance-segmentation, and detection.