Full architecture overhaul: bidirectional HCL2 ↔ JSON pipeline with typed rule classes

[kkozik-amplify]

Summary

Replaces the old DictTransformer (Lark parse tree → Python dict) with a two-stage pipeline that routes through a typed intermediate representation:

Forward:  HCL2 Text → Lark Parse Tree → LarkElement Tree → Python Dict/JSON
Reverse:  Python Dict/JSON → LarkElement Tree → Lark Tree → HCL2 Text

This enables full round-trip fidelity: parse HCL2 to JSON and reconstruct it back to semantically equivalent HCL2.

What changed

• LarkElement rule classes (hcl2/rules/) — one typed class per grammar rule with property accessors, serialize(), and _children_layout annotations
• RuleTransformer (hcl2/transformer.py) — maps Lark parse tree nodes to LarkElement tree
• Deserializer (hcl2/deserializer.py) — Python dict → LarkElement tree (reverse of serialize)
• HCLReconstructor (hcl2/reconstructor.py) — LarkElement tree → Lark tree → HCL2 text with correct spacing
• Formatter (hcl2/formatter.py) — vertical alignment and whitespace normalization on LarkElement trees
• Public API (hcl2/api.py) — load/loads, dump/dumps following the json module convention, plus intermediate-stage functions
• CLI (cli/) — hcl2tojson and jsontohcl2 entry points with flags mapping to option dataclasses
• SerializationOptions / DeserializerOptions / FormatterOptions — dataclass-based configuration for all pipeline stages
• Tests — reorganized into test/unit/ (direct rule instantiation) and test/integration/ (golden-file round-trip suites)

Highlights

• Every grammar rule has a corresponding typed Python class with child accessors
• Block vs object semantics preserved via __is_block__ markers
• Heredoc round-tripping with optional trim support
• Operator precedence parenthesization (optional force_parens mode)
• Comment preservation through with_comments serialization option
• Scientific notation, splat expressions, for-expressions all round-trip correctly
• 594 tests passing and 99% coverage

Breaking changes

• hcl2.load() / hcl2.loads() now accept keyword-only option parameters instead of positional booleans
• Old DictTransformer removed — use hcl2.transform() for the LarkElement tree or hcl2.load() for dicts
• hcl2tojson CLI flags changed to --flag style (e.g. --with-meta, --wrap-objects)