If many threads use the same Isolate (or many Isolates) and then
terminate, their PerIsolateThreadData objects are never cleaned
up, resulting in a slow memory leak and, worse, the
PerIsolateThreadData chain getting larger and larger, adversely
affecting performance.

In this situation, embedders will now be encouraged to apply
DiscardThreadSpecificMetadata against any Isolate a thread is
done with, especially if the thread is about to terminate.

Note that it is harmless to run DiscardThreadSpecificMetadata
against an Isolate for which a thread has no thread data and
per-Isolate thread data can be reestablished if a thread starts
using an Isolate again after running DiscardThreadSpecificMetadata
against it.

It is, however, an embedder error to run
DiscardThreadSpecificMetadata against an Isolate in thread with a
Locker for the Isolate in the stack or against an Entered Isolate.

This change cannot cause any change in behavior in existing apps
as the only added coded can only be reached via the new
DiscardThreadSpecificMetadata method.

R=Jakob, jochen
BUG=

Review URL: https://codereview.chromium.org/1522703002

Cr-Commit-Position: refs/heads/master@{#32909}