services/horizon: Improve performance of path finding endpoint

[tamirms]

The path finding endpoints have had increased latency and CPU consumption because of the growth in trading pairs coming from both the orderbook and liquidity pools.

This is the tracking issue for implementing performance improvements to the in-memory graph implementation.

Improvements:

• Replace DFS algorithm with DP algorithm (exp/orderbook: Improve path finding algorithm #4096)
• Represent assets in orderbook graph as int32 instead of strings (exp/orderbook: Represent assets in orderbook graph as int32 instead of strings #4102)
• Use uint128 library for liquidity pool math instead of math/big (exp/orderbook: Use uint128s for pool calculations to improve performance #4104)
• Reduce path search allocations (xdr and exp/orderbook: Reduce path search allocations #4105)

For future reference, these Ideas did not work:

• Use https://github.com/ncw/gmp as a replacement for math/big (it actually made performance worse)