Fix C/C++ frontend defects: sizeof operands, C++ extensions, catch handlers, AST cache key

[prabhu]

C/C++ frontend fixes found while comparing c2cpg with a fully resolving C++ front end. One commit per fix, each with regression tests.

Frontend (c2cpg)

• sizeof and friends. sizeof(char *) was emitted as sizeof(char): the type operand lost its abstract declarator. alignof, typeid, typeof and sizeof... were all emitted as <operator>.sizeOf, so passes that read sizeof as a byte size also read alignments, type ids and pack counts. Each now has its own operator (<operator>.alignOf, <operator>.typeId, <operator>.typeOf, <operator>.parameterPackSize), the GNU unary forms (__alignof__ x, noexcept(...), &&label) no longer fall through to <operator>.unknown, and the new operators carry no-flow semantics like sizeof.
• File extensions. Only .c, .cc and .cpp were translation units, so .cxx and .c++ sources were never parsed, and .hh headers were parsed as C. C++ sources now include .cxx, .c++ and .C; C++ headers include .hxx, .h++, .H, .ipp, .inl and .tcc; longer extensions also match in upper case.
• C++ catch handlers. Every catch body was a direct child of the try with the same order, the exception parameter was dropped (a use of e resolved to nothing), and function-try-blocks lost their handlers. Handlers now sit in one block coded catch, each at its own position with its exception bound as a local, and a function-try-block wraps the function body in a try. The CFG reads a catch group's children as alternatives, which also fixes javasrc2cpg, where a second catch clause ran after the first instead of being entered from the try body. The uninitialised-read rule treats the new binding as initialised by the throw.
• Macro handler. MacroHandler.scala held raw NUL bytes, so git treated it as binary. The expansion under a macro's INLINED call reused argument index 1, colliding with the first argument copy; it now takes the index after the last argument.
• --no-include-auto-discovery was accepted but had no action.

AST cache

• The cache key ignored the parser options. atom's header mode parses a project without function bodies and leaves its cache in <input>/.chen; a later full run over the same directory replayed those ASTs, so its graph had no method bodies and its usage slices were <empty> placeholders (cdxgen's C evidence step followed by evinse hits exactly this). Function bodies, inactive code, comments, image locations, trivial expressions, the C++ standard, include discovery and the project index now take part in the key.
• Shared tables leaked across runs. Types, member layouts and constants registered by an earlier run in the same process were lent to the next graph as extra TYPE nodes; C2Cpg.createCpg now starts from empty tables. This also fixes CWarmRestoreTests failing whenever the AST cache tests ran first in the same JVM.

Docs

The c2cpg lesson listed MACRO_DEF/MACRO_EXPANSION nodes that do not exist, a --enable-ast-cache flag that is not accepted and a frontend-args syntax atom does not parse. It now documents the parsed extensions, the pass list, the macro census flags and how macro invocations are represented. The CDT notes say the vendored core bundle ships in both CDT 12.4 and 12.5.

Testing

• c2cpg/test 1211, x2cpg/test 118, dataflowengineoss/test 53, javasrc2cpg/test 573: all pass.
• Memory-safety corpus unchanged: medium recall 62.77% / precision 98.33%, low 75.53% / 95.95%, no findings on negative controls.
• sbt scalafmtAll applied.