Synchronize concurrent include hashing in Parser

[CharlesXu-HQ]

Parser shares its include-hash cache and cycle-detection set across calls. Native threads parsing the same tracked header can race on those containers and report a circular include for an acyclic header. This happens before Runtime reaches its in-memory cache.

Guard both public hashing entry points with one recursive mutex, since recursive include traversal calls between them. The lock covers parsing only; NVCC compilation is outside it. Add an eight-thread regression that starts concurrent parses of the same tracked header and checks every digest.

Validation:

• On the base revision, an eight-thread reproducer fails with circular include may occur: shared/header.cuh for an acyclic header. With this change, it passes 10 consecutive rounds.
• The full python tests/test_cuda.py suite passes on an RTX 5090 with CUDA 13.0 and PyTorch 2.9.1+cu130, including the new concurrent parser test, header self-containment checks, and eight-process cache publication tests.
• git diff --check passes.

This change synchronizes calls through Parser's hashing methods. Direct concurrent mutation of its public containers remains unsupported; the separate in-memory cache synchronization is addressed by #12.