feat: add auralization pipeline for inference at arbitrary positions

[yaswanth169]

Summary

Adds an inference and auralization module that enables rendering Room Impulse Responses (RIRs) at arbitrary source-listener positions using trained AV-DAR models, with optional audio convolution for spatial auralization.

Motivation

The current codebase provides train.py and evaluate.py but no way to use a trained model for its core purpose — rendering audio at novel source-listener positions. This PR fills that gap with a clean, self-contained auralization pipeline.

New Files (no existing files modified)

File	Lines	Description
auralize.py	186	CLI entry point (alongside train.py / evaluate.py)
avdar/auralization/init.py	2	Package init
avdar/auralization/audio.py	38	Audio I/O and convolution utilities
avdar/auralization/engine.py	108	InferenceEngine wrapping RirRenderer
avdar/auralization/README.md	155	Feature documentation

Usage

# Render a single RIR
python auralize.py \
    --config_dir ./outputs/HAA-Classroom-16K/run \
    --source_xyz 1.0 2.0 0.5 \
    --listener_xyz 3.0 1.5 0.5 \
    --output_dir ./auralized_output

# Auralize a dry audio file
python auralize.py \
    --config_dir ./outputs/HAA-Classroom-16K/run \
    --source_xyz 1.0 2.0 0.5 \
    --listener_xyz 3.0 1.5 0.5 \
    --input_audio ./dry_speech.wav \
    --output_dir ./auralized_output

# Batch render from JSON
python auralize.py \
    --config_dir ./outputs/HAA-Classroom-16K/run \
    --positions_file ./positions.json \
    --output_dir ./auralized_output