Refactor mixed builds to only take one pass

This large refactoring has both immense performance implications and
improves mixed builds complexity / usability. Summary:

1. Queueing calls to Bazel is done in a new mutator instead of a full
   soong_build pass. Normal soong_build flow is interrupted (via a
   functional hook in blueprint) to invoke bazel and parse its response.
2. Implementing mixed build support for additional modules is as simple
   as implementing MixedBuildsBuildable. In this interface, define the
   request that must be queued to Bazel, and then subsequently define
   how to handle the returned bazel cquery metadata.
3. Mixed builds consists of only a single pass. This greatly
   improves mixed build performance.

Result:
  A 33% runtime improvement on soong analysis phase with mixed builds.

Caveats:
  C++ BazelHandler handling still remains a bit of a mess; I did what
  I could within this CL's scope, but this may require additional cleanup.

Test: Treehugger
Test: Verified that aosp_arm ninja file is bit-for-bit identical with or
without this change.

Change-Id: I412d9c94d429105f4ebfafc84100d546069e6621