Code review comment for lp://staging/~jameinel/bzr/1.17-gc-single-mem

This patch finishes up the work I started with "commit" and memory consumption.

After this patch, when doing 'bzr commit' in a --2a format repository, we no longer hold more than one copy of the file text in memory.

So --2a now only holds:
  1 file text
  2 copies of the compressed bytes

For committing a single large file with lots of small lines, this is:

  bzr.dev (last week) 457324 KB 6.164s
  bzr.dev (_add_text) 215996 KB 3.980s
  this code 123904 KB 3.863s

So this code at least isn't slower, and it cuts the memory for a large file commit by a good amount.

https://bugs.edge.launchpad.net/bzr/+bug/109114

I also checked, and I couldn't see any major change in the time for 'bzr pack', so it doesn't seem to make anything slower, while decreasing peak memory consumption by quite a bit.

(Note that the memory consumed during 'bzr pack' is unaffected, and is still considerably higher (536MiB). I'm guessing it is dominated by the 'create_delta_index' which peaks at something like 2x bytes while building and re-shaping the index.)