fix policy normalization migration keys

[MahimanaGIT]

What this does

Noticed a critical bug in the policy migration which introduced severe regressions in policy performance.
When the camera name of the robot contains _, line 112 feature_name = ".".join(parts[:-1]).replace("_", ".") in migrate_policy_normalization.py used to convert it every _ to .. Due to which the normalize_processor while normalizing the input images won't be able to find the camera key and return the tensor without normalization at line 307 in normalize_processor.py during policy execution.

Fix:

1. Added _ handling when there are multiple parts to the feature name in migrate_policy_normalization.py.
2. Updated normalize_processor.py to raise error if policy is missing keys and updated its test.

How it was tested

• Updated test_quantile_with_missing_stats in test_normalize_processor.py

How to checkout & try? (for the reviewer)

Provide a simple way for the reviewer to try out your changes.

• pytest test/test_normalize_processor.py
• Test the policy migration for a robot with camera having underscore _ in it.

[MahimanaGIT]

@AdilZouitine can you be the primary reviewer of this MR as you were the lead on normalization migration and missing key handling in normalization processing.