improvements to wav2aug

[gfdb]

This PR brings significant performance improvements to wav2aug augmentations, aligning behavior with SpeechBrain's implementations while dramatically reducing compute time.

Key Changes

speed_perturb

• Use integer percentages (90, 100, 110) instead of float multipliers (0.9, 1.0, 1.1) and then round. This ensures good GCD with sample rates, making the sinc resampling filter much smaller
• Cache torchaudio.transforms.Resample objects via lru_cache to avoid recomputing filter kernels

chunk_swap

• Replaced nested Python loops with fully vectorized gather/scatter operations
• Eliminated per-sample iteration entirely

NoiseLoader

• New preload mode
• New class that preloads all noise files into CPU RAM at initialization
• Configurable storage_dtype (default float16) for memory efficiency w extremely tiny perf. degradation
• Noise sampling becomes a fast tensor slice with zero I/O
• Memory: ~650MB for pointsource_noises pack

Other improvements
freq_drop: Ported SpeechBrain's notch filter implementation for correctness
rand_amp_clip: Fixed normalization and uses single clip value per batch (matches SpeechBrain)
time_dropout: Vectorized implementation
Wav2Aug: Simplified interface, uses NoiseLoader by default