Reduce peak memory and repeated analysis in encode/decode pipelines

[audexdev]

Summary

The current encode/decode path buffers large inputs and outputs in multiple places, and the encoder repeats expensive LPC/Rice analysis while evaluating stereo and zero-run choices.

Affected areas

• src/io/wav_io.cpp
• src/main.cpp
• src/codec/lac/encoder.cpp
• src/codec/lac/decoder.cpp
• src/codec/block/encoder.cpp

Problem details

Known memory/performance pressure points:

• WAV reader loads the full data chunk into raw, then expands into channel vectors.
• CLI loads the full .lac file before decode.
• Decoder allocates complete output vectors up front.
• Encoder stores every encoded block before copying to the final bitstream.
• Auto stereo and zero-run selection repeat block analysis and residual estimation.
• Block encoder stores residual vectors for many predictor candidates.

Acceptance criteria

• Add benchmark or measurement notes for peak memory and runtime.
• Avoid the most obvious duplicate allocations/copies in hot paths.
• Consider span-based partition analysis and moving only the winning residual candidate.
• Document whether streaming encode/decode is in scope for the current format.