refactor: Extract bytecode lowering into dedicated pass

[zharinov]

Summary

• Move bytecode constraint handling from IR generation (scope.rs) to dedicated lower.rs pass
• Fix panics for post_effects > 7 and neg_fields > 7 by adding cascade logic
• Simplify emit_epsilon and rename emit_match_with_cascade → emit_match

Why

Clean separation of concerns: IR is now unconstrained during generation and optimization passes. Bytecode encoding limits (3-bit effect counts, 5-bit successor counts) are handled in the final lowering pass before layout.

This enables future IR optimizations (like branch prefix stealing) without premature bytecode constraints.