fix(tf): fix argcheck when compressing a model converted from other backends (#4331)

When the model is converted from other backends, the input script only
contains the `model` section. This PR sets the default for any necessary
argument.

<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->

## Summary by CodeRabbit

- **New Features**
- Enhanced the data structure for model compression by adding default
keys for training steps and learning rate.
  
- **Bug Fixes**
- Improved error handling with more informative runtime exceptions for
missing training scripts.

<!-- end of auto-generated comment: release notes by coderabbit.ai -->

Co-authored-by: Your Name <[email protected]>

Author: njzjz

deepmd/tf/entrypoints/compress.py

@@ -147,6 +147,8 @@ def compress(
         10 * step,
         int(frequency),
     ]
+    jdata.setdefault("training", {"numb_steps": 0})
+    jdata.setdefault("learning_rate", {})
     jdata["training"]["save_ckpt"] = os.path.join("model-compression", "model.ckpt")
     jdata = update_deepmd_input(jdata)
     jdata = normalize(jdata)