Read a change's file list without its patch, and never cut a diff

Read a change's file list without its patch, and never cut a diff

The review tree, the file count and the landing summary were all parsed out of
one `git diff` capped at 512 KiB. When a change carries a large file the cut
lands inside it, and every file named after the cut is missing from the tree:
reviewing #143, a 546 KB deletion hid the 296 files that came after it while
the tab still said "6 changed". The cap existed to stop a page rendering every
hunk, but the pages it guarded already render one file at a time.

The inventory and the patch are separate reads now. DiffInventory runs
`git diff --raw --numstat -z`, which names every file with its status and line
counts in one walk of the trees and generates no hunks; the changed-file tree,
the counts and the landing summary read that. The Files tab then asks for the
one file it shows with DiffPaths, and the commit page keeps its per-file
ShowPatchPath. Nothing is capped by bytes: a page is as large as the file it
shows, and the inventory is the whole change however large it is.

Compare still renders every file of a range at once, and is uncapped too, so a
large comparison is one large page. Making that per-file is the next step if it
turns out to matter.

TestDiffInventoryCountsEveryFile covers the statuses, counts, renames and binary
files the inventory has to report; TestRequestFilesListsEveryChangedFile is the
regression for the tree that lost its files, and the commit page test now
asserts the end of a large patch instead of its cut.