We wish to reduce AssemblyProgram's memory usage on instruction storage.
This requires changing Disassembler*::Disassemble() to call ParseFile()
in 2 passes: (1) count bytes needed (so we can preallocate) (2) store
bytes. This CL implements the 2-pass system, focusing on callers. Details:
- Add AssemblyProgram::GenerateInstructions() as main entrance point
  from Disassembler*. It takes InstructionGenerator, which is just
  Disassembler*::ParseFile() wrapped in a Closure.
- Change Disassembler*::ParseFile() and downstream functions to const,
  to assert absence of Disassembler state change for the passes.
  - Exception: For ELF, we need to stort abs32_locations_ and
    re32_locations_ to obtain different traversal orders. This doesn't
    add/remove elements. So we make these containers mutable.
- Emit*() functions are moved from AssemblyProgram to new interface
  InstructionReceptor. Updating all Emit*() calls to accommodate.
- Add InstructionReceptor interface with 2 implementations:
  - InstructionCountReceptor for pass 1 (no-op in this CL).
  - InstructionStoreReceptor for pass 2 (calls AssemblyProgram::Emit*()
    in this CL).
- Rename Emit*Instruction() to Emit*().

TEST=Correctness without behavioral change; Courgette-gen and Courgette-apply are not slowed down, and do not use more memory.
BUG=660980

Review-Url: https://codereview.chromium.org/2462993003
Cr-Commit-Position: refs/heads/master@{#431053}