[FEA] Explore lto fragments in cudf-polars

libcudf now offers facilities to pre-bake "LTO fragments" via jit or similar and then load and use them at runtime.

In cudf-polars we have an increasing number of custom AST transformations that are hand-coded. We might imagine that turning these into LTO fragments (given we know what types might be needed ahead-of-time) would be performance-beneficial