Rephrase backend dialect definition (#950)

Summary:
Pull Request resolved: #950

As titled, rephrase it to make it clear.

Reviewed By: kirklandsign

Differential Revision: D50328853

fbshipit-source-id: 808bbbbe368448ebd47b2690dbd8d150be2ff2ab

Author: larryliu0820

docs/source/compiler-backend-dialect.md

@@ -1,7 +1,7 @@
 # Backend Dialect
 ## Overview
 
-_Backend dialect_ is a special variant of [edge dialect](./ir-exir.md), because it contains target specific nodes and metadata, after optional target specific graph transformations. In practice, a graph in backend dialect may contain operators or delegated lowered modules (see [delegate doc](./compiler-delegate-and-partitioner.md)) that are only meaningful to the target backend. One use case is that if we want to fuse operators into a single operator, for example, fusing consecutive addmm + relu to a single operator addmm_relu, we can do that here.
+_Backend dialect_ is a special variant of [edge dialect](./ir-exir.md), because it contains backend specific nodes and metadata, after backend specific graph transformations. Backend dialect is an optional stage, only needed if we want to introduce backend-awareness into the graph. More specifically, a graph in backend dialect may contain operators or delegated lowered modules (see [delegate doc](./compiler-delegate-and-partitioner.md)) that are only meaningful to the target backend. One use case is that if we want to fuse operators into a single operator, for example, fusing consecutive addmm + relu to a single operator addmm_relu, we can do that here.
 
 This document describes how to introduce backend specific operators.