[codegen, bug]: Callee kernel name generation is incorrect

Kernel call sites assume that the name of the generated function is identical to the LoopKernel's name. However, the actual name of the non-entrypoint kernel is generated during linearization as

https://github.com/inducer/loopy/blob/f67b65ccb7b377430986cf59db0747e51dfe84e7/loopy/schedule/device_mapping.py#L35-L40

---

Two ways of fixing this:
1. For non-entrypoint kernels we should always emit `CallKernel` with name equal to its `LoopKernel`, OR,
2. While emitting the code for a `Call` expression node, query the translation unit to get its `CallKernel`.

I feel (1) is cleaner(+easier to implement) and seems like it doesn't make any assumption that might bite us in the future. Opinions?