修复cosyvoice2的flow合成前后爆音

CausalMaskedDiffWithXvec forward() 开头修改为：

    token = batch['speech_token'].to(device)
    token_len = batch['speech_token_len'].to(device)
    feat = batch['speech_feat'].to(device)
    feat_len = batch['speech_feat_len'].to(device)
    embedding = batch['embedding'].to(device)

    token_len = torch.minimum(token_len, feat_len.int() // self.token_mel_ratio)
    feat_len = (token_len * self.token_mel_ratio).int()
    token = token[:, :token_len.max()].contiguous()
    feat = feat[:, :feat_len.max(), :].contiguous()
    assert feat.size(1) == token.size(1) *  self.token_mel_ratio, f"feat: {feat.size()}, token: {token.size()}"