Fix: two ring-buffer allocator defects in pto_ring_buffer.h

[chenshengxin2026]

Summary

• Bug 1 (heap wrap-around): try_bump_heap used tail > alloc_size (strict greater-than), causing deadlock when tail == alloc_size — exactly enough space exists at [0, alloc_size) but the condition incorrectly rejected it. Fixed to tail >= alloc_size.
• Bug 2 (DepListPool sentinel collision): alloc() computed idx = top % capacity, which returns 0 (the NULL sentinel slot) when top is a multiple of capacity. Fixed to idx = ((top - 1) % (capacity - 1)) + 1 so the index always stays in [1, capacity-1]. Overflow check tightened from used >= capacity to used >= capacity - 1 to match the reduced usable range.
• Both fixes applied to all three affected runtimes: a2a3/tensormap_and_ringbuffer, a2a3/aicpu_build_graph, a5/tensormap_and_ringbuffer.

Testing

• All 113 Python unit tests pass (pytest tests -m "not requires_hardware")
• Simulation tests (./ci.sh -p a2a3sim)
• Hardware tests (ctest -R test_ring_buffer once test infrastructure is available)

Fixes #429

[gemini-code-assist]

Code Review

This pull request modifies the ring buffer and dependency pool logic across several files, adjusting allocation boundary checks and index calculations. The feedback identifies a significant risk regarding the use of signed 32-bit integers for monotonic counters; specifically, signed overflow is undefined behavior and the modulo operator can return negative results for negative operands, potentially leading to out-of-bounds memory access. It is recommended to use unsigned arithmetic or explicit casting to ensure safe index generation.