Move the implementation of the rewriting engine to Langkit_Support and
add bindings to it in the generic API. Rework the C bindings to be
implemented on top of the generic API.
Note that no new testcase is added: the generic API is tested indirectly
by all existing rewriting testcases by the virtue of being the
common implementation behind all the other rewriting APIs. The Ada API
testcase will transition to the generic API once the Ada API is removed.
Symbolization is a costly operation: do it only when requested. Rework
the C API to make this possible. For now, do it automatically in the
Java API to avoid breaking existing Java programs: this will fix the
performance situation in the other C-based APIs.
Also rework lkt_compile and lkt_build_and_run test drivers so that the
language spec configuration bits are stored in the testcase's test.yaml
file directly: no need to add glue code in various places (test driver,
lkt_compile.py/lkt_build_and_run.py scripts) to support a new kind of
language spec configuration.
Whether unparsers are generated or not has implications on legality
checks for the grammar, so it makes sense for this to be a property of
the grammar itself: introduce an Lkt annotation as well as a
langkit.parsers.Grammar argument for it, and remove the corresponding
CompileCtx argument.
Also rename "generate_unparser" to "generate_unparsers" for consistency
across the codebase.
Transition all relevant testcases to it. In the process, port
docstrings/comments from test.py files to expected_concrete_syntax.lkt
(now test.lkt) sources and fix stylechecks issues there.
This makes the convention consistent with the DSL, and avoids
workarounds for conflicts with Lkt keywords: in Libadalang, the Null
token can stay Null, instead of null_tok (no API breakage needed).
