Fix HTTP performance regression

[reidsunderland]

This fixes the performance regression in the HTTP transfer class.

Instead of calling build_opener (which is very slow) for every __open__, it is only done in connect now.

I also made another change, which removes the setting of self.connected = False in the __open__ method. With that there, it forces a "re-connect" (and build_opener call) every time __open__ was called. Now connect gets called only when check_is_connected returns False, which will mainly happen if there is a change to the "sendTo" URL (if it was downloading from google.com, then we switch to downloading from example.com, it will re-connect).

If we have a feed that downloads from many different domains/sources, we would still see the performance issue, because of the re-connecting. I left a note in the comments saying that I think build_opener could actually only be done even less frequently, in __init__. But I was worried that that was too large of a change to make.

This change is just one commit, so it should be easily cherry-pickable if we want to make a 3.00.55post2 release

[github-actions]

Test Results

249 tests   248 ✅  1m 29s ⏱️
  1 suites    1 💤
  1 files      0 ❌

Results for commit 0fe61aa.