Optimize device lib linking

[pxl-th]

• Cache contents of devlib to avoid reading it every time there is linking.
• Do not link devlib if it is not needed.
  This applies only to ("hc", "hip", "irif", "ockl", "opencl", "ocml") devlibs, as other seem to be always required.

Before (Julia 1.10):

julia> @time sin.(x);
 11.552189 seconds (30.58 M allocations: 1.394 GiB, 1.27% gc time, 59.74% compilation time: 2% of which was recompilation)

After (Julia 1.10):

julia> @time sin.(x);
  7.220562 seconds (13.40 M allocations: 878.159 MiB, 1.49% gc time, 93.13% compilation time: 2% of which was recompilation)

Also test suite is now 2x faster to run 20 min -> 10 min.