[ir] [refactor] Slim ir.h and reduce the build time by ~4%

[xumingkuan]

Related issue = #689

Hopefully, this will also reduce the build time.

Benchmark (after commit fe533e5, on my laptop with Windows 10 and i7-7700hq CPU, rebuild 3 times before and after this PR):
before: build time: 2m59s, 2m51s, 2m49s.
after: build time: 2m44s, 2m44s, 2m43s.
before (reproduce)(rebuild after writing this PR and checking out to the previous version): build time: 2m50s, 2m51s, 2m53s.

After merging with master (after commit 936cd97):
before: 2m52s, 2m53s.
after: 2m47s, 2m47s.

After merging with master again (after commit fbc12ab):
before: 2m55s.
after: 2m46s.

[Click here for the format server]

[xumingkuan]

Looks like I can't move template classes' functions to ir.cpp. (for example, member functions of LaneAttribute<T>)

[xumingkuan]

And any other template functions, like template <typename T> inline void StmtFieldManager::operator()(const char *key, T &&value).

[yuanming-hu]

Right, template functions cannot be moved to cpp unless you do explicit template instantiation. I would suggest simply leaving these in the header file.

[yuanming-hu]

Great, thanks! The IR part is more organized now. Please fix Travis, and I'll merge this in.

4% compilation time improvement is awesome! I appreciate that you have done a solid benchmark. Note that not every file includes ir.h so the actual compilation time improvements for code related to the compiler is, I guess, probably ~10%.

I believe you'll get more performance boost after splitting Expression (frontend) from chi (middle-end). We can do that as well sometime in the future.