Inline AncestralSequence::mlCharProbAt and AncestralSequence::mlCharProbAtF methods

[althonos]

Hi Ari!

I was exploring the PRANK code and doing some profiling to see if I could make it any faster. After profiling with Valgrind, I found that the AncestralSequence::mlCharProbAt was in a very hot path: on my example of 12 sequences, it was being called 3M times, with a total cycle count of 150M cycles, out of a total execution of 400M cycles.

However, all this function is doing is loading data from memory, so it's not really normal to take 50 cycles per fetch; besides, hiding it in a function prevents the CPU from pipelining load instructions when used inside of a loop. By inlining this one function, the same invokation of PRANK went from 400M cycles to 250M cycles.

In terms of timing, that's how it improved on my machine:

[althonos]

Hi @ariloytynoja, I've seen you updated the repo recently, would you consider merging this PR as well?