CPU compile (#691)

* build and load shared object for cpu compile

* nits

* cpu compile tests pass

* cpu compile tests pass

* fix preamble for g++

* donation

* fix gpu buffer donation

* reuse prebuilt libraries

* faster contiguity conditoins

* fix test

* rid compiler warning

* fast erf

* Fix float16 for compile and add more types to cpu compile

* Remove a forgotten comment

* use cached libs

* nits

---------

Co-authored-by: Angelos Katharopoulos <a_katharopoulos@apple.com>

mlx/compile.cpp

@@ -319,6 +319,9 @@ void compile_simplify(
       case 1:
         v = *a.data<uint8_t>();
         break;
+      case 2:
+        v = *a.data<uint16_t>();
+        break;
       case 4:
         v = *a.data<uint32_t>();
         break;