It's redundant and never used, because shared code in some sitations assumes
NativeTrampolineCall to be same size as NativeCall. However, they are different
sizes (3 and 5 words on ARMv6), so first one is padded with nops up to second.

Presence of NativeTrampolineCall::instruction_size encourage it's use and thus
introducing bugs. Example is nativeCall_before:

inline NativeCall* nativeCall_before(address return_address) {
  if (NativeTrampolineCall::is_at(return_address - NativeCall::instruction_size)) {
    ...
  }
  ...
}

It will be incorrect if `NativeTrampolineCall::instruction_size` used here, but
looking good.

NativeCall::instruction_size is not C++ constant and computed at VM start after
determing weither it's running on ARMv6 or ARMv7.
Making NativeTrampolineCall::instruction_size to be always equal
NativeCall::instruction_size implies making one not constant too and adding
relevant initializing code.