Fix "Fouble" (Fake Double) precision errors across TFLM Kernels

Problem Statement

Multiple kernels perform intermediate scaling calculations using 32-bit float precision before assigning to a 64-bit double variable. This "Fouble" (Fake Double) pattern leads to precision loss and state drift in recurrent models.

// Example values that require more than 24 bits of mantissa (float precision)
float input_scale = 1.2345678f;
float feature_scale = 1.2345678f;

// (BAD) Fouble: Multiplication happens in float, then is cast to double
double fouble_result = static_cast<double>(input_scale * feature_scale);

// (GOOD) Defoubled: At least one operand is cast to double first
double defoubled_result = static_cast<double>(input_scale) * feature_scale;

This issue has been identified and addressed in several previous PRs (e.g., #1492). However, those attempts were ultimately discarded because the fix requires a synchronized change within the main TFLite codebase.

Proposed Changes

Force double-precision arithmetic by casting an operand early (e.g., static_cast(a) * b).

Hard Constraints & Coordination:

• Bit-Exactness: All changes must maintain parity with TFLite reference kernels. Coordination with the main TensorFlow repository is required.

• Vendor Kernels: Optimized implementations for specific hardware (e.g., ARM, Cadence) must be updated in tandem with reference changes to ensure consistent behavior across platforms.

• Rollout Strategy: We will avoid a single monolithic PR. Instead, we will submit incremental PRs grouped by source file to facilitate easier regression testing and binary-size monitoring.

Target Files:

• kernels/activations_common.cc
• kernels/div.cc
• kernels/elementwise.cc
• kernels/hard_swish_common.cc
• kernels/leaky_relu_common.cc
• kernels/softmax_common.cc
• kernels/squared_difference.cc
• kernels/sub_common.cc
• kernels/svdf_common.cc