bannertopdf: close output stream so the generated PDF is not truncated

[mtaplits]

Fixes #212.

Problem

cfFilterBannerToPDF() opens a stream on the output file descriptor with fdopen(outputfd, "w"), but on the success path it never closes it — fclose(outputfp) only appears on the error paths. The stream therefore keeps its last buffer, and cfFilterChain() closes the raw output fd and exits immediately after the filter function returns, so that tail is discarded.

The lost bytes are the cross-reference table and the %%EOF marker. With the universal filter the next filter in the chain, pdftopdf (backed by PDFio), cannot parse the file and exits with status 1, which kills the chain:

cfFilterChain: pdftopdf completed with status 1.
cfFilterChain: pdftopdf (PID …) stopped with status 1
ERROR: universal filter failed.

This breaks printing the CUPS test page (application/vnd.cups-pdf-banner) and banner pages (job-sheets=standard). Ordinary documents are unaffected because their chains do not start with bannertopdf.

It is a regression from 1.x: there bannertopdf was a standalone filter binary writing to stdout, so normal process exit flushed the stream. The 2.x rewrite made it an in-process library function invoked by cfFilterChain()/cfFilterUniversal(), where the caller closes the fd and exits without giving the filter a chance to flush.

Verification on master

Running the universal filter on the test page template, before and after this change (builds of this branch, same configuration):

env CONTENT_TYPE=application/vnd.cups-pdf-banner CUPS_DATADIR=/usr/share/cups \
    CUPS_SERVERBIN=/usr/lib/cups CUPS_SERVERROOT=/etc/cups PPD=<queue>.ppd \
    CHARSET=utf-8 FINAL_CONTENT_TYPE=application/vnd.cups-raster \
    /usr/lib/cups/filter/universal 1 "$USER" testprint 1 "" \
    /usr/share/cups/data/testprint > out.raster

	exit	bytes	chain
before	1	4	pdftopdf stopped with status 1
after	0	485766	all filters status 0

bannertopdf run standalone produces 49736 bytes for the same input; in the chain pdftopdf was receiving 49152 = 12 × 4096, i.e. truncated exactly on a stdio buffer boundary, with the missing 584 bytes being the xref table and %%EOF.

Content after the fix is intact: the page renders with the same ink density as the template scaled to the same canvas (0.64% vs 2.55% for default-testpage.pdf, ratio 0.252 against a 4×6-to-A4 area ratio of 0.248).

Notes

• make check: 7 pass, 4 fail — testpdf2, test-analyze, test-pdf, test-ps, all with Font no was not loaded, i.e. the missing ttf-dejavu checkdep. These fail identically without this change and are unrelated to it.
• The test suite does not currently exercise cfFilterBannerToPDF() at all, which is why this regression was not caught. I have not added a test here since it would need a new fixture (a PDF-io-based assertion on the generated stream); happy to add one if you would like it in this PR.
• The related but separate truncation in cfPDFPagesFP() (cupsfilters/pdf.c, missing fflush() before pdfioFileOpen()) is already fixed upstream by e14bad4061 / Fix printing error: pdfio output "Missing Root object." #167 and is not touched here.