feat(dsl): C ABI and Python DSL checking, closing p7-s5 (#43)

[JArmandoAnaya]

The last slice of p7-s5: scena::dsl::check_* reaches all three of Scena's
surfaces, so the DSL checker is not a C++-and-CLI-only feature.

C ABI

scn_check_dsl_file / scn_check_dsl_string produce an opaque scn_dsl_check
carrying the diagnostics, the type count and the file count. Checking needs no
engine — it is a frontend service — which is why the result is its own handle
rather than an engine's diagnostic list, and why its borrowed strings stay valid
until scn_dsl_check_destroy instead of until the next mutating call.

Two decisions worth naming:

• A failing check still produces a handle. That is exactly the case whose
  diagnostics you want. Only a NULL argument returns without one, which is also
  how a caller tells a null argument apart from an unreadable path — both answer
  SCN_ERROR_INVALID_ARGUMENT, per the Status model's host-misuse split.
• A zero-initialized scn_dsl_check_options turns the standard library off.
  A zero struct has to mean zero for an appendable ABI struct, so NULL options
  is what means "the C++ defaults, library included". Both are documented on the
  declarations.

capi/ links scena::frontend-dsl PRIVATE, like the XML frontend; nothing in
capi.h names a C++ type.

No ABI version bump. The header's rule is that adding symbols moves the
minor, but nothing is published yet and every prior symbol addition (p5-s4,
p5-s5, p6-s1, p6-s2) also landed at 10000. Bumping only for this one would
misrepresent the others; the released version is the maintainer's call.

Python

scena.check_dsl_file / check_dsl_string return a DslCheck — status,
diagnostics, type_count, file_count, falsy unless Ok. The XML loaders
return (status, scenario) because the scenario is the payload and the findings
are secondary; a check has no payload, so a named object reads better than a
four-tuple.

python/examples/check_dsl.py runs in CI: a clean source, a defective one, and
an import resolved through a search path.

Parity audit

The audit only ever looked at scena::Engine methods, so the frontend entry
points could lag a binding unnoticed — which is exactly what this PR would have
risked. It now audits the namespace-scope entry points of scena::xml and
scena::dsl alongside the methods, same exclusions-carry-reasons rule:
xml::validate_* (loading validates on the way through) and dsl::load_* (the
lower half of check_*; its result is an AST no binding can carry).

test_parity.py's doctored-header test copies the frontend headers into its
sandbox too, so it still fails for the right reason.

Tests

• capi_test.cpp +6: coverage, counts, search paths, the standard-library
  toggle, an unreadable path, and the null-argument sweep.
• c_consumer.c: the same surface exercised from pure C11.
• test_dsl_check.py +10.
• 1310 gtest (+6), 194 pytest (+10).

Docs

scena-check.md gains the C and Python worked examples; c-api.md and
python.md link to them; the DSL coverage matrix records all four checking
surfaces as a table of its own.

Closes #43