Better safeguard on unloadable MKL (#793)

* Better safeguard on unloadable MKL

Tries to safeguard julia-actions/setup-julia#300 JuliaLang/julia#55878 where the user has requested an incorrect architecture and thus we may believe we can load MKL until we check `MKL_jll.is_available() == false`. The reason is because we're being lied to about `@static if Sys.ARCH === :x86_64 || Sys.ARCH === :i686` which then makes it take this branch and then MKL does not exist. We could instead check `isapple` another type of workaround here and 99% of users would be perfectly fine, but if this is safe this is slightly better for the 1 remaining Intel Apple user that exists in the wild somewhere.

* Update LinearSolve.jl

Author: ChrisRackauckas

src/LinearSolve.jl

@@ -52,17 +52,21 @@ const CRC = ChainRulesCore
         # MKL_jll < 2022.2 doesn't support the mixed LP64 and ILP64 interfaces that we make use of in LinearSolve
         # In particular, the `_64` APIs do not exist
         # https://www.intel.com/content/www/us/en/developer/articles/release-notes/onemkl-release-notes-2022.html
-        using MKL_jll: MKL_jll, libmkl_rt
+        using MKL_jll: MKL_jll
         const usemkl = MKL_jll.is_available() && pkgversion(MKL_jll) >= v"2022.2"
     else
-        global libmkl_rt
         const usemkl = false
     end
 else
-    global libmkl_rt
     const usemkl = false
 end
 
+@static if usemkl
+   using MKL_jll: libmkl_rt
+else
+   global libmkl_rt
+end
+
 # OpenBLAS_jll is a standard library, but allow users to disable it via preferences
 if Preferences.@load_preference("LoadOpenBLAS_JLL", true)
     using OpenBLAS_jll: OpenBLAS_jll, libopenblas