Full scheduler::Symbol support + Keyword argument forwarding

[carstenbauer]

Terminology: By "top-level" keyword arguments I refer to keyword arguments that are directly provided to the parallel function API, e.g to tmapreduce, tmap, etc. (in contrast to those provided to a Scheduler constructor).

Principle:
If scheduler isa Symbol, then (most) "top-level" keyword arguments are forwarded to the constructor of the Scheduler that corresponds to the symbol (if any). If scheduler isa Scheduler, to avoid ambiguity, we error if a scheduler option is provided as a "top-level" keyword argument.

Examples (see docstrings and runtests.jl for more):

julia> tmapreduce(sin, +, 1:100; scheduler=:static)
-0.12717101366042027

julia> tmapreduce(sin, +, 1:100; ntasks=2)
-0.12717101366041983

julia> tmapreduce(sin, +, 1:100; chunksize=10, scheduler=:static)
-0.12717101366042027

julia> tmapreduce(sin, +, 1:100; chunksize=10, scheduler=StaticScheduler())
ERROR: ArgumentError: Providing an explicit scheduler as well as direct keyword arguments (e.g. chunksize) is currently not supported.
[...]

julia> @tasks for i in 1:100
           @set ntasks=4*nthreads()
           # non-uniform work...
       end

julia> @tasks for i in 1:100
           @set begin
               scheduler=:static
               nchunks=10
           end
           # work ....
       end

More:

• StaticScheduler and DynamicScheduler now can also take a ntasks keyword argument which is just an alias for nchunks.
• DynamicScheduler now defaults to ntasks=nthreads() (again).

Besides implementing these features, I found and fixed some minor code, docstring, and testing bugs.

Closes #76.

[carstenbauer]

I think that tests are failing because of #82.

[MasonProtter]

Wow that ended up being quite involved to do! Looks good to me. This is really starting to make me wish there was a package or something that made it easier to deal with "computed kwargs", because doing this right can get quite complex.

[MasonProtter]

Oops, meant to hit "approve"