fix: preserve speaker_encoder in checkpoints to allow training resume

[haosenwang1018]

Problem

In finetuning/sft_12hz.py, the speaker_encoder weights are explicitly deleted from the state dict before saving checkpoints (lines 150-153). When training is resumed from one of these checkpoints, model.speaker_encoder is None, causing a crash on the first forward pass.

Related to #204 (bug 1: "Speaker encoder is removed from checkpoints – breaks resume")

Root Cause

drop_prefix = "speaker_encoder"
keys_to_drop = [k for k in state_dict.keys() if k.startswith(drop_prefix)]
for k in keys_to_drop:
    del state_dict[k]

Fix

Remove the speaker_encoder deletion. Checkpoints now include all model weights, allowing training to resume correctly. Users who need smaller inference-only exports can strip the speaker_encoder weights in a separate post-processing step.

Impact

• Checkpoint size: Increases by the size of the speaker_encoder weights (~small relative to the full model)
• Resume: Now works without crashes
• Backward compatible: Existing code that loads checkpoints without speaker_encoder will still work (missing keys are handled by the model's load logic)