Update on "[kernel] Add template based unboxing"

Adding a new feature to allow users to bypass codegen and register their
kernels directly. This is very useful for custom kernels for custom ops.

Example usage:
```
Tensor& my_op(RuntimeContext& ctx, const Tensor& self, const Tensor& other, Tensor& out) {
    // ...
    return out;
}

Kernel my_kernel = Kernel.make_boxed_kernel("my_ns::my_op",EXECUTORCH_FN(my_op));

register_kernels({my_kernel});
```

Differential Revision: [D51553099](https://our.internmc.facebook.com/intern/diff/D51553099)

[ghstack-poisoned]