Standardize TFLM Reference Kernels to Single Rounding Requantization

Issue

TFLM's Reference Kernels (and TFLite's one) use a legacy "Double Rounding" algorithm, resulting in an approximately ~1.8% off-by-one error rate. This diverges from the mathematically correct "Single Rounding" used by TFLite's Optimized kernels, primarily XNNPack and ruy, which serve as the "Golden Reference." Context: Single-round was introduced in TFLite by https://github.com/tensorflow/tensorflow/pull/50290 to address https://github.com/tensorflow/tensorflow/issues/25087 in 2021 but it wan't enabled by default due to concerns about regression.

This inconsistency creates a critical "Validation Trap":

• Validation: Developers validate models using TFLite in Python (correct Single Rounding).
• Deployment: Deploying to TFLM (incorrect Double Rounding) causes output mismatches, leading to unexpected accuracy drops and engineering churn.

Proposal: Standardize Reference Kernels to Single Rounding

I propose standardizing the Reference Kernel behavior to Single Rounding, ensuring TFLM matches the TFLite behavior.

Action Plan:

1. Enabling TFLITE_SINGLE_ROUNDING by default in Bazel & Makefile build. (Optionally, we can consider introducing a new switch to use the double-rounding)
2. Vendor Alignment and Coordination: Propagate this standard to optimized TFLM vendor kernels (e.g., ARM CMSIS-NN, Cadence Xtensa). Coordination is essential to ensure these paths align with the new mathematical standard, maintaining ecosystem consistency.