Fix directory detection heuristic

[cromedome]

Fixes #1781. Supersedes #1782.

Problem

_build_location walks up from the caller's directory looking for the application
root. Its test for "is this an app root?" was "does it hold both lib/ and bin/?"
— which also describes every Perl distribution root, most home directories, /usr,
and /usr/local.

This was harmless while the upward walk was effectively stalling, since it rarely
left the script's own directory. 2.1.0 made the walk work, and the weak test
immediately started selecting directories far above the application. Apps then
resolved config.yml, views/, public/ and logs/ against the wrong root, which
is the template and public-file failures in #1781.

Reproducing

The trigger is any directory holding both lib/ and bin/ above the app. A
Dancer2 checkout already has lib/:

mkdir bin
prove -l t/auto_page.t     # fails
rmdir bin
prove -l t/auto_page.t     # passes

That's why a plain cpanm Dancer2 fails for some people and not others: ~/bin +
~/lib is enough, and the walk climbs out of the extracted tarball into the home
directory. Reconstructing that:

mkdir -p /tmp/fakehome/bin /tmp/fakehome/lib
git archive HEAD --prefix=Dancer2-2.1.0/ | tar -x -C /tmp/fakehome
cd /tmp/fakehome/Dancer2-2.1.0 && prove -Ilib -r t/

11 failing test files on main, 0 on this branch.

Fix

Recognise an app root by what it actually holds, closest-first as the walk climbs:

1. a .dancer file — the explicit marker
2. environments/, views/ or public/
3. config.<ext> for any extension Config::Any handles
4. lib/ + bin/ — kept for apps with no Dancer2 artifacts, but checked last so it
   only wins when found closer than real evidence. Still skips blib/.

Because t/ holds config.yml, views/ and public/, it now matches immediately
and the walk never approaches the distribution root or ~.

A second commit stops the walk at a checkout or distribution boundary (.git,
.hg, .svn, Makefile.PL, Build.PL, dist.ini, cpanfile) rather than
escaping above it. That only applies to directories already rejected as app roots,
so a dancer2 gen app — which ships a cpanfile and a Makefile.PL alongside
.dancer and config.yml — is accepted several tiers earlier.

The walk is also canonicalised with realpath up front, and now stops when
parent() reaches its own fixed point instead of comparing against rootdir. A
relative $subdir became a chain of .. segments, which made the blib/ regex
meaningless and the root comparison non-portable. That's the real issue behind
#1782, whose stated premise — path('.')->parent being a fixed point — doesn't hold
on current Path::Tiny.

Note on the test change

t/issues/memleak/die_in_hooks.t asserts on stderr and was relying on detection
failing: it previously found no app root, fell back to its own directory, and so
never loaded t/issues/config.yml with its logger: "Note". It now has a config of
its own stating that requirement.