Fix the three translator emitters that produce all 144 generated-C warnings

[shai-almog]

The generated C is down to 3 warning shapes / 144 instances on the iOS
simulator leg, from 3,035 instances in 8 kinds when the census started. What is
left is three emitter defects, each of which is one fix that erases a whole
category:

instances	flag	where it comes from
98	-Wunused-variable	a C auto is emitted per bytecode local slot
42	-Wparentheses-equality	if ((a == b)) shape in an emitted condition
4	-Wimplicitly-unsigned-literal	a literal too large for a signed int

The unused-variable hypothesis, stated as a hypothesis.
BytecodeMethod.appendMethodC() walks localVariables and emits one C auto per
slot unconditionally, without consulting whether optimize() left a use. Every
method with a slot the optimizer killed then contributes one warning. Confirm
cheaply before writing any code: count how many of the 98 messages in the census
JSON name a locals_N_ identifier. If most do, the fix is at that emission
point.

-Wimplicitly-unsigned-literal (4) is worth reading as a possible bug rather
than a style nit -- it appears in java_lang_Long.m and
com_codename1_sensors_GestureEngine.m, and a literal silently reinterpreted as
unsigned changes comparison results.

Do not suppress these. Blanket-silencing translator output is how a codegen
defect hides forever; the whole point of putting generated C in scope was to fix
the emitters. If a residual genuinely cannot be fixed, the narrow fallback is a
#pragma GCC diagnostic prologue emitted from ByteCodeClass.generateCCode()
-- one emission point, visible in the generated file, and it behaves the same
under Xcode clang, clang-cl and gcc. Use the GCC spelling: gcc does not
understand #pragma clang diagnostic, and -Wunknown-pragmas is inside
-Wall. Every such pragma keeps a matching baseline row, because the pragma
silences the compiler and the row is what a human has to justify.

Acceptance: the generated group reaches 0 instances in the census summary,
and its baseline rows are deleted (stale rows are a hard failure by design).

Found by the native warning census (scripts/check-native-warnings.py).