Write out the scan, instead of a box saying the picture was unreadable

[tannevaled]

reader v0.4.2 → v0.5.0. No code change.

What was wrong

picture() hands anything that is not a JPEG or a JPEG 2000 to samples(), and
samples() refuses data that cannot fill the page before it reads any of it:

func stencil(data []byte, w, h int) (image.Image, bool) {
	stride := (w + 7) / 8
	if len(data) < stride*h {
		return nil, false          // <- a fax always landed here
	}

A /CCITTFaxDecode image arrived still compressed, so there were never enough
bytes, so picture returned false and figure emitted a placeholder. Every
scanned picture in a document came out as a box saying it could not be read.

That guard is the reason this was a missing figure rather than a wrong one — the
same data painted without a length check is noise, which is what render v0.10.0
had to stop doing.

Measured

Two binaries calling latex.Source, each verified with go version -m:

old  github.com/go-pdfkit/reader v0.4.2   pdffont v0.3.0   extract v0.1.0
new  github.com/go-pdfkit/reader v0.5.0   pdffont v0.3.0   extract v0.1.0

Corpus: all 2 268 forms in /Users/Shared/pdfforms. Prevalence is over the
1 633 real documents.

	files
compared	2 268
source differs	59
… of the 1 633 real documents	57 (3.5%)
… of the 635 vendor fixtures	2
longer after	57
shorter after	0
status changed	0

Across those 57 real documents:

	before	after
\framebox{...unreadable image}	291	50
\includegraphics	12	253

241 placeholders became figures, one for one. Verbatim, from
fr-cerfa/cerfa_10983.pdf:

before:  \framebox[41.0pt]{\rule{0pt}{21.0pt}unreadable image}
after :  \includegraphics[width=41.0pt,height=21.0pt]{image000.png}

and from fr-cerfa/cerfa_10701.pdf, two on one page:

before:  \framebox[31.0pt]{\rule{0pt}{16.0pt}unreadable image}
after :  \includegraphics[width=31.0pt,height=16.0pt]{image001.png}
before:  \framebox[31.0pt]{\rule{0pt}{16.0pt}unreadable image}
after :  \includegraphics[width=31.0pt,height=16.0pt]{image003.png}

The 50 that are still unreadable are not faxes. They are colour spaces
samples() does not claim to read without a full renderer behind it, and they
were not part of this bump.

Not measured

Wall clock and peak memory: three other corpus jobs were on the machine, so a
timing number would be fiction. The output diff does not depend on load.

Gates

go vet, gofmt, exact 100% statement coverage, nine cross-compile targets, all
pass locally with GOWORK=off CGO_ENABLED=0.