[Mangler] Accelerate the variable-length encoding by using native integers.

Previously we were manipulating APInts every time we went down a level on the
huffman tree. This commit changes the search to use uint64_t. We copy the last
few bits of the APInt into the uint64_t and manipulate that value, which is much
much faster because we don't need to copy the APInt (which is typically a few
hundred bits long). The uint64_t can't overflow because it is not possible to
encode ascii in a huffman tree and use sequences that are 64bits long.

Author: nadavrot