fix(summary): support sparse tensor inputs in memory accounting

[DangTAnh]

Summary

Models that take sparse tensors (e.g. GCN adjacency matrices) crashed summary() with NotImplementedError: Cannot access storage of SparseTensorImpl, so no summary — including the Param # column from #150 — was produced. Sparse layouts expose no single storage, so input-memory accounting now sums the member tensors (indices + values) instead.

Changes

• torchinfo/torchinfo.py: new get_tensor_memory helper with sparse fallback; get_total_memory_used uses it.
• tests/torchinfo_test.py: new test_sparse_input regression test (GCN with sparse COO adj).

Testing

• pytest tests/torchinfo_test.py tests/exceptions_test.py tests/mapping_test.py --no-output — 94 passed.
• Repro from Summary doesn't give the number of parameters for a GCN #150 (GCN + sparse adj): crashed before, now prints Param # correctly (172/172 params); verified COO, CSR, CSC, and dict inputs.
• 3 golden-file errors in the full run are pre-existing torch-version drift (fail identically on clean tree).

Related

Fixes #150

[TylerYep]

Thanks!