Superfluous downstream checks in Makefile parallelism

[wlandau-lilly]

In make(..., parallelism = "Makefile"), drake bulids outdated targets and then checks all the downstream targets to see if they are also outdated. This snowball effect is often unnecessary, and it is not present in any of the other parallel backends.

load_basic_example()
make(my_plan)
clean(large)
make(my_plan, parallelism = "Makefile", jobs = 8)
## ...
## check 1 item: large
## target large
## Rscript -e 'drake::mk(target = drake::as_drake_filename("report.md"), cache_path = "C:/projects/.drake")'
## Rscript -e 'drake::mk(target = "regression1_large", cache_path = "C:/projects/.drake")'
## Rscript -e 'drake::mk(target = "regression2_large", cache_path = "C:/projects/.drake")'
## load 1 item: large
## check 1 item: regression1_large
## load 3 items: coef_regression2_small, large, small
## Rscript -e 'drake::mk(target = "summ_regression1_large", cache_path = "C:/projects/.drake")'
## check 1 item: 'report.md'
## Rscript -e 'drake::mk(target = "coef_regression1_large", cache_path = "C:/projects/.drake")'
## load 1 item: large
## check 1 item: regression2_large
## Rscript -e 'drake::mk(target = "summ_regression2_large", cache_path = "C:/projects/.drake")'
## Rscript -e 'drake::mk(target = "coef_regression2_large", cache_path = "C:/projects/.drake")'
## load 1 item: regression1_large
## check 1 item: summ_regression1_large
## load 1 item: regression1_large
## check 1 item: coef_regression1_large
## load 1 item: regression2_large
## check 1 item: summ_regression2_large
## load 1 item: regression2_large
## check 1 item: coef_regression2_large

Standaline Makefiles do not seem to have this problem.

all: c

c: b
	touch -a c

b: a
	touch -a b

a:
	touch -a a

clean:
	rm -f a b c

make
## touch -a a
## touch -a b
## touch -a c
make
## make: Nothing to be done for `all`.

And when I modify a by hand in a text editor, c does not build.

make
## touch -a b

[wlandau-lilly]

Fixed, thanks to #168. Added a new next_stage() function inspired by parallel_stages(). FYI: @kendonB, outdated() should be faster now too. It just uses next_stage() to grab the first outdated targets and then assumes everything downstream is also out of date. In other words, all the downstream hashing and checking is skipped.