Faster indexing math in a few kernels (#1589)

* wip: faster compiled kernels

* faster general unary with uint specialization

* index type in compiled, unary, binary, ternary, copy

* fix jit

* jit fix

* specialize gather + scatter

* nit in docs

mlx/backend/common/compiled_cpu.cpp

@@ -279,7 +279,7 @@ void Compiled::eval_cpu(
 
   // Figure out which kernel we are using
   auto& shape = outputs[0].shape();
-  bool contiguous = compiled_check_contiguity(inputs, shape);
+  auto contiguous = compiled_check_contiguity(inputs, shape);
 
   // Handle all broadcasting and collect function input arguments
   std::vector<void*> args;