Reduce the C++ compilation duration

[k-ye]

Concisely describe the proposed feature

This is derived from #2196 and #2200. We should aim to reduce the time spent on compilation/linkage. Note that the compilation duration refers to the time to build the Taichi C++ codebase, not to JIT compile from Python to backends.

Describe the solution you'd like (if any)

A few possible action items:

• Profile the compilation duration and figure out the major bottleneck.
• If necessary, clean up the #include header dependencies.
• Set up a continuous compilation benchmark to detect regression.

Additional comments

• This isn't super urgent. Filing this one so that we don't forget.
• That said, we should have a quantitative way to decide when this is hurting the dev experience too much and needs to be prioritized.