Fix #53; replace beam search with greedy sampling

[nkoppel]

Fixes #53. Changes all mutexes/rwlocks to their blocking std counterparts and uses tokio::task::block_in_place so we can call UnboundedReceiver::blocking_recv in an async context.

Furthermore, replaces the beam search of LlamaSession::start_completing with a call to LlamaSession::start_completing_with with a greedy sampler and max_predictions as the number of unused tokens in context. I did this because I'm fairly sure that the beam search wasn't working correctly and was not updating LlamaSessionInner.tokens correctly.