[refactor] Reduce compilation time by splitting and shortening headers

[archibate]

Concisely describe the proposed feature
Currently compliling taichi takes too long time can be a great resistance for contributors especially when your work involving common headers like ir.h or lang_util.h (I guess @xumingkuan suffers this most).

Describe the solution you'd like (if any)

1. reduce length of these common headers by removing unused functions and classes.
2. move some of the inline functions in .h to .c, and then .h only holds function declaration.
3. split a large header like ir.h into multiple headers, since only part of the functions are used.
4. some common utils may use macros instead of inline functions to reduce c++ lexer's pain.

Additional comments
It's slowing down recently. Not sure what added to these headers caused this.

TODO: add compile-time test/comparison result here.

Update: I also found that clang-8 is 5x slower than clang-7.

[xumingkuan]

Thanks for proposing this! I think we can try to remove this as the first task:

taichi/taichi/ir/ir.h

Lines 1555 to 1556 in d4a9c18

	#include "taichi/ir/statements.h"
	#include "taichi/ir/visitors.h"

It's virtually increasing the length of ir.h.

(Or maybe all files that include ir.h need statement.h?)

[k-ye]

Covered in #2203