This PR addresses open-mmlab#1843.

Instead of calling inspect.stack() to read the entire stack and its
associated source files from disk, walk up the stack to get only the
specific frame that we need (see [1] for additional information).

This makes imports in downstream projects ~2.5x faster in my local dev
environment. For mmaction2, for example:

Before:

    $ python -m timeit -n1 -r1 "from mmaction.apis import init_recognizer, inference_recognizer"
    1 loop, best of 1: 1.94 sec per loop

After:

    $ python -m timeit -n1 -r1 "from mmaction.apis import init_recognizer, inference_recognizer"
    1 loop, best of 1: 754 msec per loop

[1] https://stackoverflow.com/a/42636264/895769

Explain why we avoid `inspect.stack()` with link to PR