[wsttiger]

Adds a sparse-access tutorial comparing matrix-data QROM oracles with arithmetic banded oracles using independent fixed-point matrix references. Both routes execute block, complex-action, scratch-restoration, Walk, and QSVT checks; a four-point banded case verifies clipped boundaries. Metadata covers slots, alpha, workspace, alternate data forms, and Hermitian dilation.

Extends example 07 with an executed 2x2 sparse SPD inverse at three angle bits, fresh degree-83 CKS phases, independent ideal/quantized NumPy solves, and layered error budgets. The original five-dimensional dense solve remains and now also uses analytically known scale/sign corrections instead of fitting a scalar. Binomial-tail coefficients avoid overflow, and both cases verify their complex response with an independent signal model.

Validation on CUDA-Q 0.16.0.post1, Python 3.12.3, qpp-cpu:

• Sparse introduction passed: maximum 16 executed qubits, 23.87 s in the initial direct run.
• Extended inverse passed: maximum 14 executed qubits, 11.86 s in the initial direct run; the sparse case took 9.90 s.
• python3 -m pytest tests/python/test_examples.py -k 'sparse_block_encoding or 07_matrix_inversion_qsvt' -q: 2 passed in 79.82 s.
• Sphinx HTML build with -W --keep-going: passed independently.
• YAPF 0.43.0 and whitespace checks passed.

Independent plan 05 PR against main; requires no heat application, sibling example, library change, or new dependency. Full-state real-response extraction is explicitly simulator postprocessing.