as_symmetric: deduplicate repeated imports of the same tensor

[mawad-amd]

Problem

Calling as_symmetric() twice on the same external tensor (or on tensors backed by the same physical allocation) repeats all the work with no deduplication:

1. DMA-BUF export/import/map — full export_dmabuf_handle → mem_import_from_shareable_handle → mem_map → mem_set_access cycle each call
2. VA space consumption — each call bumps current_offset, burning heap VA space for a redundant mapping to the same physical memory
3. Collective peer refresh — refresh_peer_access() fires every time, triggering TCP FD exchange + barriers across all ranks

Suggested fix

Add a lookup in VMemAllocator.import_external_tensor() keyed on alloc_base (from get_address_range). If the same physical allocation was already imported, return a view into the existing VA mapping at the correct offset instead of creating a new one.

Should also batch refresh_peer_access() — if the same segments are already mapped on peers, skip the re-exchange.

References

• iris/allocators/vmem_allocator.py:220-296 — import_external_tensor
• iris/symmetric_heap.py:276-298 — as_symmetric + unconditional refresh_peer_access()