[Bug] Fix DLPack field offset truncation and two allocator bugs.

[duburcqa]

Issue: N/A

Brief Summary

copilot:summary

Three host-side memory-safety fixes, found while root-causing a Genesis simulation NaN that appeared once a SNode tree grew past 2 GiB:

• DLPack export truncated in-tree field byte offsets to signed int (field_to_dlpack). For a field placed past 2^31 bytes in its tree, the wrapped offset failed the >= 0 guard and was silently dropped, so the exported view aliased the tree base: reads returned unrelated fields and writes corrupted them. Fixed by keeping the offset size_t/uint64_t end to end (DLTensor::byte_offset is uint64_t). Regression test: test_dlpack_field_offset_past_2gib (marked slow + run_in_serial, 2.24 GB allocation), verified to fail before the fix and pass after.
• UnifiedAllocator overstated the chunk tail (tail = data + size + allocation_size instead of data + allocation_size), so a follow-up allocation could be handed an address past the real end of the mmap'd chunk. Regression test added in host_memory_pool_test.cpp.
• CachingAllocator::allocate dereferenced an erased iterator (it_blk->second after mem_blocks_.erase(it_blk)); erase order swapped.

Walkthrough

copilot:walkthrough

[github-actions]

Total: 5 file(s) changed, +30 -5 code lines.

[github-actions]

Diff coverage: 25% · 16 lines, 12 missing

[github-actions]

Total: 5 file(s) changed, +30 -5 code lines.

[github-actions]

Diff coverage: 25% · 16 lines, 12 missing