Optimize context copies for parallel loops

[dsharlet]

When executing code in parallel, each thread needs its own copy of the eval_context object, to avoid race conditions. For large programs, where the symbol table can have many thousands of entries, this added a lot of overhead.

This PR fixes a few performance issues related to this:

• Adds an is_closure flag to let_stmt, which allows us to communicate to evaluate exactly which symbols we need to copy to the thread local context, instead of copying the entire thing.
• Adds uninitialized_allocator and uses it for the symbol table. This way, when we allocate memory for the context, we don't need to memset it via std::vector's constructor.
• Move the user configurable functions from eval_context to a separate object eval_config, and only store a pointer to that.

Here's an example of what programs look like now with closures:

out.y = loop(parallel, [buffer_min(out, 1), buffer_max(out, 1)], 1) {
 closure {in, out, out.y} in {
  out.out.y = crop_dim(out, 1, [out.y, out.y]) {
   intm.out.y = crop_dim(out, 1, [out.y, out.y]) {
    out.x = loop(parallel, [buffer_min(out, 0), buffer_max(out, 0)], 1) {
     closure {in, out, out.x, out.out.y, intm.out.y} in {
      out.out.x = crop_dim(out.out.y, 0, [out.x, out.x]) {
       intm.out.x = crop_dim(intm.out.y, 0, [out.x, buffer_max(out.out.x, 0)]) {
        call(<anonymous target>, {in}, {intm.out.x})
       }
       call(<anonymous target>, {out}, {out.out.x})
      }
     }
    }
   }
  }
 }
}

[vksnk]

This is a very cool idea!