[XrdPfc] Backstop crash from negative dir count in purge check

[h2zh]

A negative root directory count, converted to size_t, makes
vector::reserve() throw std::length_error and abort the daemon from
the purge thread (xrootd#2808). Clamp n_calc_dirs to at least
1; reserve() is only a capacity hint and fill_pshot_vec_children()
walks the actual tree, so purging proceeds correctly.

This is only a backstop. The count drifts negative because the
accounting is asymmetric: directory removals are counted whenever the
empty-dir sweep unlinks a dir and erases its DirState node, but
directory creations are only counted for cache-miss opens
(!m_existing_file). A file written to disk out-of-band (e.g. the
Pelican cache self-test file) re-creates the DirState nodes without
ever being counted. The real fix should be counting m_NDirectoriesCreated
whenever a file open has to create DirState nodes in memory,
regardless of m_existing_file. (m_NFilesCreated must stay gated on
!m_existing_file to avoid over-counting cache hits.) Having said that,
upstream XRootD developers may have more to consider.

[matyasselmeci]

I built xrootd-5.9.6-1.2 with this patch; I haven't put it into osg-testing yet because we're having odd VMU test failures (that aren't related to this patch, but I'd like to get some clean tests before putting the new build into osg-testing).