Use class unloading in dex2oat for verify and extract

Unload the main classloader in between each dex file compilation to
reduce RAM. This frees the whole java heap and associated linear
allocs. This is only used for quickening since filters that do
compilation may require loaded classes in the compiler and oat
writer.

This reduces dex2oat peak PSS for compiling a large app from 196MB
to 135MB.

Only works for verify and extract since the current approach is
incompatible with oat writer patching. b/63911263

Added a verification override that reads the compiled class status
to avoid ever verifying classes that were quickened (since this
is not supported and causes failures).

There is still some duplicated verification for some class with
superclasses in other dex files.

Support for quicken will be added in a follow up CL.

Bug: 63467744
Test: test-art-host
Test: test/testrunner/testrunner.py --interpreter --host -j40
Change-Id: Id0e4f84eb5db91d6143f752b498f4832a5b25b6e
diff --git a/runtime/class_table.h b/runtime/class_table.h
index 8616dfb..a259725 100644
--- a/runtime/class_table.h
+++ b/runtime/class_table.h
@@ -250,6 +250,12 @@
       REQUIRES(!lock_)
       REQUIRES_SHARED(Locks::mutator_lock_);
 
+  // Filter strong roots (other than classes themselves).
+  template <typename Filter>
+  void RemoveStrongRoots(const Filter& filter)
+      REQUIRES(!lock_)
+      REQUIRES_SHARED(Locks::mutator_lock_);
+
   ReaderWriterMutex& GetLock() {
     return lock_;
   }