PTX: Conditional reference-to-value rewrite

We should not unconditionally rewrite all parameters to be passed by-value, as there's actually a limit:

> `__global__` function parameters are passed to the device via constant memory and are limited to 32,764 bytes starting with Volta, and 4 KB on older architectures.

Not doing the rewrite requires that CUDA.jl allocates a box on the GPU that contains the argument. This also needs to be a stream-ordered allocation, in order to respect the kernel's lifetime.

The issue names no files or tests. Start by tracing where GPUCompiler.jl unconditionally rewrites kernel parameters to by-value arguments, then inspect the CUDA.jl path for boxing and stream-ordered allocation; done means large parameter arguments avoid the rewrite while respecting the kernel lifetime.

Written by the indexing model from the issue text.