fix(core): Address memory consumption and enhance FFT reliability

- Refactored NetworkDataProvider to use a chunked streaming and decoding model, resolving excessive memory usage. Memory consumption for a 15-minute audio file is reduced from ~1GB to ~20MB.
- Unified the FFT implementations in MathHelper for Scalar, SSE, and AVX paths to ensure consistent behavior. The SIMD paths are now toggleable via static properties for easier debugging.
- Enhanced accuracy of FFT operations through improved window function generation, cosine approximations, and proper IFFT scaling.
- Optimized core FFT operations including bit reversal and complex multiplication for improved performance.

Author: LSXPrime