epubveri for Sigil 0.3.2

• A book's documents are read once each rather than once per finding.
  Sigil puts the cursor on a finding using an absolute character offset, so
  the plugin computes one for every row. Each computation re-read the whole
  document and rebuilt its line table: a cost proportional to the file being
  paid once per finding, which is the wrong way round — a long document with
  many findings paid for its length over and over.

  Measured on the worst book of the reference shelf, 6 859 findings:
  epubveri validates it in 0.52 s and the plugin was then spending
  1.88 s turning the answer into rows. The wait was 2.40 s, of which 78%
  was the plugin re-reading files it had already read. It is now 0.53 s,
  and the row-building part is 0.011 s.

  Nothing about the output changes — same rows, same order, same offsets,
  verified line by line against the old computation. The table is built on
  demand, one entry per file, and thrown away with the report rather than kept
  in a module global: run is called again for the next book, and a global
  would hold offsets for a working directory that has since been deleted and
  rebuilt — stale by exactly the amount the user had edited.

  The test asserts the property rather than the timing — one read per file,
  however many findings it has — because a timing test on a fast machine
  passes whatever the code does. It was checked by reverting the change and
  watching it report 40 reads for 40 findings in one file.