Numeric constant-pool lowering

The CUDA bytecode analyzer and emitter support numeric ldc, ldc_w, and ldc2_w instructions. When a constant pool is available, integer, long, float, and double entries are resolved and emitted as PTX immediates.

This admits integer literals beyond sipush range as well as literal long, float, and double values. Non-numeric constant-pool entries remain rejected. The CP-free lowering entry point remains conservative and rejects these opcodes because it cannot resolve their values safely.

Regression coverage includes EligibleLdc* fixtures and ptxas_round_trip_ldc_constants.