The only purpose of the binary is to show local Shellfu path and it's
in every single Shellfu script.  Let's save people some typing.

Since __SHELLFU_DIR__ is not known at install time, its value is purely
"theoretical" and we need to be able to override it.

There's no real reason to have it public.

These variables are not specific to shellfu internals but in all ways
specific to pretty.sh module.  The prefix is mostly historical cruft.

Let's clean that up.

We needed to improve TF_RUN for this to work properly:

 *  fixup wrapping so that shebang, which influences
    detection of SHELLFU_COMPAT, comes first (if there is one),

 *  and call the test script directly for the shebang to matter
    (i.e. do not force `bash` in Bash mode upon all).

Submodules were never implemented, but also never missed; it's pointless
to keep the test stubs.

Although normally we prefer to sat "result must look like oracle",
that is, have result on RHS, and oracle on LHS of a comparison, with
udiff this produces far more readable result, e.g.:

    --- oracle/try_import_fn.new
    +++ result/try_import_fn.new
    -foo

if foo was expected but is missing,

    --- oracle/try_import_fn.new
    +++ result/try_import_fn.new
    +bar

if bar came unexpected.

so that tfkit can now be tracked in separate repo.  (In future we might
add it as submodule)

Once simple.sh has been cleaned up and simplified in previous ~10
commits, it turns out that all it does is manage subtests.

Reflect that in the name.

(Doc will be updated later.)

Support test folder structure where test, oracle and generated result
data are stored each in separate directory.  This makes it easier to
ensure legible diff output without cluttering the filenames too much.

...so that it can be understood by future reader.

This new hierarchy makes it easier to work simultaneously on data and
the script, simplifies deployment to tempdir and makes the launching
script naming consistent with simple.sh.

As a followup to last revamp:

 *  Use more extensible test definition format,
 *  separate testing framework from setup script and split to includes,
 *  update test scripts as well as data,
 *  consolidate test output,
 *  and make some small improvements/fixes on the way.

 *  Enable running all tests via make,
 *  simplify tests (lay off data/tmpdir manipulation to framework),
 *  set rules for naming, exit statuses, etc,
 *  remove beakerlib dependency.

 *  Split into functions as appropriate;
 *  add docstrings;
 *  reduce redundancy;
 *  run everything in C locale (=> update oracles),
 *  add subtest boundaries to output.