Centralize from_ir conversion; switch IR to frozen dataclasses

[twizmwazin]

Summary

• Move all JPype-touching from_ir factory logic out of the sootir/ package and into soot_manager.py as private _convert_* functions. The sootir/ package is now a pure data layer with no Java interop.
• Switch every IR dataclass from unsafe_hash=True to frozen=True, catching post-init mutation at runtime.
• Tighten types in soot_manager.py (return type annotations on the conversion helpers, narrower union types for _build_instance_invoke and the binop dispatch table, _JavaObj alias for JPype-returned objects).

Stacked on top of #41 — please merge that first.

Why frozen is now possible

SootPhiExpr previously had to be constructed without block indices and then mutated post-init in a second pass:

phi_expr = SootValue.IREXPR_TO_EXPR[ir_expr]
phi_expr.values = values  # post-init mutation

The new code threads a per-method _Ctx (containing stmt_to_block_idx) through _convert_value, so SootPhiExpr is built in a single shot with its final (value, block_idx) tuples. No mutation, no global side-channel.

Notes

• Conversion functions are now module-private (_convert_*) since they are only used by run_soot. No public API change in sootir/.
• The IR dataclasses remain hashable — frozen=True enables value-based __hash__ automatically, so the existing pattern of using SootBlock as a dict key in basic_cfg / exceptional_preds still works.

Test plan

• Ruff passes
• pytest tests/ still passes (needs JPype runtime, can't run in dev sandbox)
• Spot-check that SootMethod.block_by_label still resolves, that frozendict lookups for basic_cfg / exceptional_preds still work
• Verify Phi expressions are constructed correctly on a shimple-formatted JAR

🤖 Generated with Claude Code

[angr-bot]

Corpus decompilation diffs can be found at angr/dec-snapshots@master...angr/pysoot_42