Optimizing .dex IR strings lookup

Looking up strings is a very common operation during bytecode
instrumention (since we keep a set of unique strings even allocating
a string requires a lookup first)

This change is the result of a couple of days of evaluating various options:
- std::map
- std::unordered_map
- std::set
- insertion sort
- custom lookup index

The std::unordered_map has superb runtime performance but at a high
memory cost (~5Mb for large.dex). Crafting a specialized hash table
using the same hash function (DBJ2a) resulted in a slightly faster index
using only ~30% of the std::unordered_map<const char, String*> memory.

Using this new string lookup, the overall time for stress_entry_hook over
medium.dex improved considerably: before: 1050.294 ms, after: 264.143ms

Test: bazel test :dexter_tests

Change-Id: I9b5cb889707a498846c8f2530b28f58564a0598d