Fix compression_config dropped when loading compressed models from JSON config

[SuperMarioYL]

Summary

parse_json_full_config() parses the compression_config section into a local
comp_config, but the returned FullConfig(...) omits the compression_config=
argument. Because FullConfig.compression_config defaults to None, every config
loaded through this function comes back without any compression metadata,
silently discarding the parsing work.

This is the path used when reloading a previously compressed model:

# angelslim/engine.py (prepare_model)
slim_config = parse_json_full_config(slim_config_path)   # reads angelslim_config.json
...
self.slim_model.from_pretrained(
    ...,
    compress_config=slim_config.compression_config,      # always None before this fix
)

So the compressed-model save → load round-trip loses the compression
configuration. The YAML parser (SlimConfigParser._get_configs) already passes
compression_config correctly — this change brings the JSON parser in line with it.

Changes

• Pass compression_config=comp_config to FullConfig(...) in parse_json_full_config.
• Remove a verbatim-duplicated calibrate/transform parsing block in the same
  function (the second copy recomputed identical values).
• Raise a descriptive ValueError for a missing required JSON section
  (model_config / compression_config) instead of a bare KeyError, matching
  the validation style of the YAML parser.
• Add CPU-only regression tests (tests/test_config_parser.py) covering the
  round-trip and the missing-section path. They need neither a GPU nor model
  weights.

Testing

python -m pytest tests/test_config_parser.py -q

pre-commit run --all-files passes (black / isort / flake8) on the changed files.