Optimize Base64Decode function

[Tina-1300]

This PR replaces the existing Base64Decode implementation with a more optimized version :

• Pre-allocates the output string to avoid dynamic reallocations.
• Decodes 4 Base64 characters into 3 bytes per iteration.
• Branchless main loop for better CPU performance.
• Handles padding '=' efficiently after the main loop.
• Uses a static lookup table initialized once, reducing overhead on repeated calls.

Performance improvements are noticeable for large Base64 inputs.