If we're given a device name of an LVM internal layer, we shouldn't
just ignore the layer name and snapshot something else. Instead,
assume something's gone horribly wrong and fail hard.

Also tweak some diagnostics and comments, and add new tests for this
and the not-a-LV scenario.

On several systems, udevadm has been observed not to give output for
some LVs for no terribly obvious reason, causing rsbackup-snapshot-hook
to refuse to work with those LVs (#72).

dmsetup also lets us canonicalise device names to work around the
original issue which caused udevadm to be invoked (#23), but is closer
to the source of truth, and thus more likely to be reliable than
udevadm.

As a side effect of this rework, VG names containing hyphens should
now be handled more convincingly.

And fix the errors thus brought to light (all in the test suite,
fortunately).

This probably explains why the test suite didn't pick up 7b84f914
back in 2015.

(While I'm here, add a couple of --must-run to t-hook-fsck-fail that
ought to have been there.)

This is a (currently undocumented) feature that allows a backup
to be excluded from consideration as a link destination, for
instance if it is no longer trusted for some reason.

/usr/share/doc/rsbackup/ was ending up with both rsbackup-manual.html
and rsbackup-manual.in.html, which was untidy.

fixed #91

fixes #101

re #105

fix #105. Unclear why things work/failed where they did.

fixes #106

fixes #98

Quadratic behavior reduced to n.log n. Better is surely available but
this will do for now.

(Distinguish from Render::Context...)

Not actually very useful given the reports we actually generate.

...since that's what all the production callers use anyway.

re #63

re #63

re #63

re #63

This reverts commit 1c99194c.

--noatime turns out to be a specific to a small set of Debian
releases. It is replaced by --open-noatime. When that is more
generally available I'll start using that.

fixes #101

#92

re #100

fixes #97

fixes #100

- RSBACKUP_TIME="" is equivalent to unset
- tests.sh:s now has only one execution line

re #100

fixes #96

Includes a test for parseFloat.

fixes #99

See travis branch for attempts to get them working again.