(un)specify associativity of reduce and mapreduce

[stevengj]

Currently, the reduce and mapreduce implementations are left-associative, but this is not documented in the manual. For accuracy reasons (#4039) and possibly for future performance reasons (in parallel implementations), left-associativity is not necessarily the best choice. My suggestion is:

• Explicitly document that the associativity of these functions is implementation-dependent and unspecified.
• Provide reduce_left and mapreduce_left functions that are documented to be left-associative, for applications that need this (e.g. Monads.jl).

cc: @StefanKarpinski, @pao

[pao]

Perhaps skip providing a foldl equivalent higher-order function; you rightly point out that in Julia code this is probably more clearly and performantly expressed as a loop.

[stevengj]

After confirming with @JeffBezanson that the intention was indeed for reduce to have unspecified associativity, I have patched the documentation to clarify this.