split reduce op into multiple libraries, accelerate the compiling

[dzhwinter]

fix #10874 #9306

• split reduce_op into 5 libraries, so our machine can compile the libraries in parallel mode
  what's more, separate the functor into different source files can reduce the compiled templated code in exponential order.
• change the .cu to .cu.cc to accelerating the compiling.
  why: [Speed up compiling]: reduce the NVCC compiling (some .cu operators can be compiled by G++) #5491

before split reduce op into multiple operators

real	16m35.665s
user	15m56.777s
sys	0m50.414s

after

[dzhwinter]

The reduce op will have performance problems because it uses a lot of broadcast, which is extremely slow. And the template needs a cleanup.

[reyoung]

I think we might rewrite this function by not using template generation too much. Splitting them into files just a temporary solution.

[dzhwinter]

Yes. I can't agree with you more. I have clarified that the template needs a cleanup and the broadcast issues, so we need to rewrite the reduce op. Just leave that job when we have more spare time.

[reyoung]

Cool