fix: preserve speaker_encoder in checkpoints and improve sample rate error

[haosenwang1018]

Problem

Two issues in the finetuning script (sft_12hz.py / dataset.py):

1. speaker_encoder deleted from checkpoints — resume impossible

Lines 150-153 of sft_12hz.py explicitly delete speaker_encoder weights before saving:

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]

When resuming from a checkpoint, model.speaker_encoder = None → crash on the next forward pass.

2. Unhelpful sample rate error message

dataset.py:105 asserts sr == 24000 but doesn't show the detected sample rate or suggest a fix, leaving users confused.

Fix

1. Remove the speaker_encoder deletion — keep weights in checkpoints so resume works. A comment explains the rationale.
2. Improve the assertion to show the actual sample rate and suggest the ffmpeg conversion command.

Fixes #204