Clear exc.__traceback__ after handled exceptions to prevent memory leak

[baconcheese113]

Summary

• When wrap_app_handling_exceptions catches and successfully handles an exception, exc.__traceback__ retains a reference to the frame's locals, which include the ASGI send/receive closures and per-request state
• If the exception object is long-lived (e.g., a module-level HTTPException reused across requests — common in auth libraries), this pins per-request objects in memory permanently. GC cannot collect these because the module holds a strong non-cyclic reference to the exception
• The fix adds exc.__traceback__ = None after the handler runs and the response is sent. Unhandled exceptions (raise exc) are not affected — their traceback is preserved for propagation

Reference chain that causes the leak

exception.__traceback__
  → frame (wrapped_app's except block)
    → locals:
      → sender (closure) → captured `send` → RequestResponseCycle (~19 KB)
      → receive → RequestResponseCycle

Precedent

• CPython #88863 — same bug class in asyncio.open_connection. Fixed by clearing references after exception handling. Backported to 3.10/3.11.
• Uvicorn PR Use pep562 as a dependency rather than bundling? #1604 — fixed a related cyclic reference leak by nulling self.parser after connection loss.

Test plan

• Reproduced the bug on unpatched code (shared exception's __traceback__ is not None after handling)
• Confirmed the fix resolves it (__traceback__ is None after handling)
• Verified traceback doesn't accumulate across 100 requests
• Verified unhandled exceptions (no handler found → raise exc) still preserve their traceback
• New test test_handled_exception_clears_traceback passes on both asyncio and trio
• Full test suite: 925 passed, 2 xfailed, 0 failures