tp: binder_tracker: fix performance issue due to insert/erase heavy workload + FlatHashMap

On a 10 minute trace with ~500k binder_transaction events, ~50% of the
trace parsing time is spent in binder_tracker due to a performance issue
in FlatHashMap. The map implementation never cleans tombstones outside
of a resize, so a heavy insert+erase workload degrades to full scans of
the hash map's capacity (since it becomes 100% full with tombstones).

For now, fix by switching to std::unordered_map. Myself and Primiano
will look into improving the hash table itself next week.

Change-Id: Iedca0e490e4a024fb59b84b485be250bd579f9d3