Integrated baseline (Habets/PRA) into the evaluation pipeline

move waveform2spect from collate_fn (cpu) to trainer (gpu)

Model evaluation compared to synthetic RIRs

analyze data distributions

minor changes

add T60 percentage [Eval]

LSD metric is based on reverbed speech and not clean rir

in eval script, select multiple for each metric

improvements to full evaluation flow

wave2spect is now parallel

new best cfg

load rir from evaluation to reference

full_model_eval creation

minor changes

add DDIM to DIff Model