[maokovski]

Sinkhorn previously captured a new CUDA graph on every long fused solve. This change caches the graph and reuses it across compatible calls.

• Refresh inputs and reset solver state before replay, then copy results into each call’s output buffers.
• Rebuild the cache when the device, stream, tensor layout, or kernel configuration changes.
• Exclude warmup from the requested iteration count.
• Propagate graph errors instead of retrying with modified state.

Add regression tests for cache reuse with new inputs, exact iteration counts, output lifetime, configuration and stream changes, and error propagation.

Changed files:

• torchmorph/optimal_transport.py
• test/test_optimal_transport.py

Validation: Python syntax and diff whitespace checks passed. Runtime tests are delegated to the server CI; CUDA execution and results remain to be confirmed. No performance benchmark results are claimed.