[bugfix] fix #5596: floating number calculation converter

[Alanscut]

Changes:

• Use floating number calculation converter to process the multiply in duration.
• Fix issue Float number calculation loses accuracy #5596.
• Any suggestions for the pull request will help me, thanks.

[coveralls]

Coverage decreased (-0.09%) to 88.437% when pulling d48376e on Alanscut:issue-5596 into 528ac2b on moment:develop.

[ichernev]

Merged in b4e0676

[ichernev]

I tweaked the original comment a bit. This multiply code basically makes the multiply a tad less wrong, if that.

[ichernev]

Reverted in 15b82f5

The floating point issues have been notice a while ago, and all this does is revert to the old code (36e5 instead of 1000 * 60 * 60). There is common knowledge that while floating point ops are not perfect, a single op (A * B in this case) is as accurate as possible. This code tries to "improve" the accuracy of simple multiplication (via the newly added multiply fn). However, this is just not possible.

In general the way to "solve" such issues is to compute a complex expression in different order (i.e reorder computations). However there isn't much that can be done in this case.

Read more about it here: #2978

If you just want to convert hours to milliseconds, the formula is hours * 36e5 == ms, and you can round that however way you like (i.e Math.round), chances are that fractional milliseconds are not that useful. But moment can't really make such decisions.