fix: SM50 error-object leak/null-deref + fence QI-failure leak (clang-tidy)

[cyyever]

Three memory-safety fixes surfaced by a clang-tidy (clang-analyzer) sweep:

• airconv/dxbc_converter.cpp — SM50ErrorInternal leaked on every successful compile. The six SM50* entry points new'd an error object at entry but only returned it via *ppError on error paths, so it (and its LLVM buffer) leaked on success. Now held in a std::unique_ptr, released to the caller only on error.
• Same file — null *ppError write. SM50Initialize treats ppError == NULL as legal (if (ppError) *ppError = nullptr;) but the error paths then wrote *ppError unconditionally, crashing such callers. All error paths now go through a set_error() helper that no-ops when ppError is null.
• d3d11/d3d11_fence.cpp — fence leaked on QueryInterface failure. CreateFence/OpenSharedFence new'd the impl (refcount 0) then returned QueryInterface(...); a QI failure orphaned it. Now held in Com<> via ref().

Each changed TU compiles clean.

[3Shain]

Please sign off your commits (assuming works are done by human)

[cyyever]

@3Shain Signed

[3Shain]

@3Shain Signed

Is it possible to use a real name?

[cyyever]

@3Shain Sure, I have resigned.

[3Shain]

Format your commit message and split into two by scope

[cyyever]

fixed

[3Shain]

fixed

Have you ever checked existing commit messages on main branch...

It should start with fix(scope):

[3Shain]

Why closed? Anyway thanks for spotting the leaks.