perf: optimise VM runtime execution loop

[Copilot]

The VM interpreter loop contained several avoidable overheads on every instruction fetch, making runtime slower than necessary.

Changes

• Remove dead division per instruction — DWORD line = vm->location / 4 + 1 was computed every loop iteration but never used.

• OPCOND_AL fast-path — _condition() (a function call + switch dispatch) was invoked unconditionally for every instruction. Since the vast majority of ARM instructions use the "always" condition (0b1110), the condition field is now checked inline first; _condition() is only called for genuinely conditional instructions. __builtin_expect hints the branch predictor accordingly.

  // Before: always calls _condition()
  if (!_condition(vm, instr >> 28)) return;
  
  // After: skips the call entirely for OPCOND_AL (the common case)
  DWORD cond = instr >> 28;
  if (__builtin_expect(cond != OPCOND_AL, 0) && !_condition(vm, cond))
      return;

• Mark internal helpers static — _condition, _calcshift, _calcimmediate, all exec_* helpers, and exec_instruction were externally visible despite being purely internal. Without static, the compiler must preserve them as addressable symbols and is conservative about inlining. This lets -O2 inline them freely.

• Fix UB in _calcimmediate() — when the rotation field is zero, the old code evaluated Imm << 32, which is undefined behaviour for a 32-bit type. Added an early return for the zero-rotation case.