Interpreter improvements: extract opcodes, arrow syntax, simplify LOAD_STRING

[fglock]

Summary

• Extract 84 opcode handlers from BytecodeInterpreter.execute() into new InlineOpcodeHandler.java, reducing method size from 12007 to 7137 bytes (under 7500 JVM JIT limit)
• Convert all switch statements to arrow syntax (case X -> {}) for clarity, eliminating break statements
• Add comments explaining the nested try/catch/finally structure (outer for cleanup, inner for Perl eval/die)
• Simplify LOAD_STRING and LOAD_BYTE_STRING opcodes — remove redundant cache-check optimization (benchmarked: no measurable difference on ExifTool CanonRaw.t)
• Add 76 missing opcode mappings to InterpretedCode disassembler
• Remove dead executeArithmetic() method (326 lines)
• Improve LOAD_BYTE_STRING javadoc in Opcodes.java
• Reformat code across the codebase
• Fix check-bytecode-size.sh for renamed package path
• Improve parser error messages for croak.t (11→44/330)
• Fix $# check to allow special array names ($#+, $#-, $#{^CAPTURE})

Test plan

• ./gradlew classes builds successfully
• bash dev/tools/check-bytecode-size.sh — all methods under limit
• ./gradlew classes testUnitParallel --parallel shadowJar — all tests pass
• Benchmarked LOAD_STRING simplification on ExifTool CanonRaw.t — no regression

Generated with Devin