As observed in
https://gitlab.com/gitlab-com/gl-infra/scalability/issues/77, a
rate-limited blob request would still cause disk I/O since
`assign_ref_vars` called FindCommit to look up the given commit.

To avoid this, we need to do a number of things:

1. Move rate-limiting to occur before the `assign_ref_vars` filter.

2. Call `extracts_path` directly to pull out the filename of the
request.

3. Use the path as the key for rate-limiting. This means that filenames
with different commits will now be rate-limited in the same bucket.
That should be okay since the intent of the rate-limiting is to block
abusive behavior, and requesting the same filename with different commit
IDs at a high rate could create performance issues.

4. Redirecting to another blob path would still cause Gitaly RPCs
because the controller would call `assign_ref_vars`. Now, just render a
plain text messsage with a 429 response code.

Closes https://gitlab.com/gitlab-com/gl-infra/scalability/issues/77