Attempt to improve speed by constant factors

[tdhock]

fixes #11

[tdhock]

the slowdown was due to two factors

• returning ParamsLoss struct copy (fixed by passing pointer arg instead)
• using unordered_map operator[ multiple times (fixed by using it once to get pointer)
  This result suggests that storing mean param as a plain struct member (not in unordered_map) may be even faster, as shown below
  

[tdhock]

storing parameters in class members instead of unordered_map results in significant speed improvements, but still not as good as old code which only supported square loss (no other distributions).

[tdhock]

omitting variance estimation for square loss results in small speed improvements, but still not as fast as old code (cv) with plain square loss implementation (no other distributions).

[tdhock]

further improvements could be obtained by omitting storage and calculation of total weighted squares cumsums and removing that term from the mean_norm loss, then adding back at the end.