Fix PCEN preprocessor and improve checkpoint config merging

[shaing10]

Summary:

• Fix LibrosaPcen: accept hop_length param (backward compat with n_mels), handle torch.Tensor input, squeeze singleton dims to prevent 5D tensor crash, update deprecated librosa.core.pcen to librosa.pcen
• Add max_freq merging from checkpoint in merge_with_checkpoint()
• Add PCEN preprocessor config (_preprocessor_with_pcen.yaml)
• Add optional max_freq field to config validator

Context

The Amber dolphin models (echo, whistle) use a Spectrogram → PCEN preprocessor chain. These models failed on inference due to:

1. LibrosaPcen.init() not accepting hop_length (the param saved in their checkpoints)
2. LibrosaPcen.call() passing a torch.Tensor to librosa.pcen (expects numpy)
3. np.expand_dims adding a redundant channel dim on top of the existing one from the data pipeline, creating a 5D tensor that crashes ResNet182D.forward()

All 4 Amber models (barks, echo, buzz, whistle) verified working on this branch with identical outputs to client/amber.

Test plan

• Barks model (4af2w6lt) — inference passes, output matches baseline
• Echo model (bki984uw) — inference passes, output matches baseline
• Buzz model (ccgojzau) — inference passes, output matches baseline
• Whistle model (g8gtuypk) — inference passes, output matches baseline
• Verify existing models (non-PCEN) still work on master (no regressions — PCEN changes are additive, n_mels param still supported)