Lua client fixes

[djungelorm]

Three unrelated defects in the Lua client, each fixed in its own commit.

Global leak when coercing an argument. The result of the pcall that coerces an argument was assigned to an undeclared ok, making it a global, so every call needing a coercion wrote to and left a value in the global namespace. Only ok is now declared local: the value it is assigned alongside is the loop variable holding the argument and has to stay that way, since declaring both would scope a new local to the branch and the coerced value is read after it — the argument would then be silently sent uncoerced. A comment records this.

Broken encoder.client_name. It padded a name to encoder.CLIENT_NAME_LENGTH, which is never defined, so the function raised an error whatever it was passed. It is a leftover from before the protocol used protocol buffers and nothing calls it, so it is removed.

Enumeration name never passed to its Lua type. set_values passed self._enum_name, which is never assigned; the field set when the type is constructed is _class_name, so the name reaching _create_enum_type was always nil. This is harmless today only because that function ignores the service and class names it is given and uses only the values, hence no changelog entry.

Testing. New test_coercion_does_not_leak_a_global in the Lua client tests passes a plain table where a list is expected, which is a coercion that succeeds, and asserts no global ok is left behind. Passing a list where a tuple is expected does not exercise this, as a tuple's Lua type is List, so the argument is already considered valid and no coercion is attempted.