Consolidate and group related test variants for ops, cumops, reductions, and shape tracker

[jafioti]

Motivation

• Reduce duplication by grouping closely related test cases for unary/binary ops, cumulative ops, reductions, and shape tracking.
• Ensure multi-variant behaviors (broadcasting, multi-axis reductions, and cumops along non-last axes) are validated in existing test harnesses.
• Keep test code easier to maintain by folding small scattered checks into existing tests.
• Add a targeted movement test for gather and inverse_permutation to cover those operators.

Description

• Merged softmax axis-1 and cumulative-op variants into existing unary tests and added reference implementations cumsum_ref_2d, cummax_ref_2d, and cumprod_ref_2d in src/hl_ops/unary.rs and moved their checks into test_softmax/test_cumsum.
• Added broadcasting variants to test_add and test_mul in src/hl_ops/binary.rs to cover more broadcast shapes.
• Folded expand/squeeze checks into the test_unsqueeze movement test and added a new test_gather_and_inverse_permutation in src/hl_ops/movement.rs.
• Consolidated multi-axis reduction checks into existing test_sum and test_mean, and kept test_prod in src/hl_ops/reduction.rs.
• Combined symbolic simplification/substitution checks into a single test in src/shape/symbolic.rs and merged the shape-tracker expand/broadcast checks into test_permute_and_expand in src/shape/tracker.rs.

Testing

• Ran cargo fmt successfully to format the changes.
• No automated unit tests were executed as part of this change (cargo test was not run).

---

Codex Task