Implement and test tensor math operations

Summary

Implement and validate `fensor` math-operation traits so non-transactional tensor math reaches v1 parity targets (subject to documented sparse/layout support boundaries and structured unsupported errors).

Scope

• Implement/complete:
  • `TensorUnary` (`exp`, `ln`, `round`)
  • `TensorMath` (`add`, `sub`, `mul`, `div`, `log`, `pow`, `rem`)
  • `TensorMathScalar` (`*_scalar`)
  • `TensorReduce` (`max`, `min`, `product`, `sum`)
  • `TensorReduceAll` (`max_all`, `min_all`, `product_all`, `sum_all`)
  • `TensorReduceBoolean` (`all`, `any`)
  • `TensorMatMul` (using `matmul_output_shape` contract)
• Enforce shape/dimension validation with clear structured errors
• Preserve explicit unsupported boundaries where required (especially sparse edge paths)

Explicit Non-Goals

• Transaction lifecycle/LSM logic in `tc-collection`
• Introducing broad new platform abstractions from `ha-ndarray` beyond what is needed for `fensor`

Acceptance Criteria

• Math op matrix passes for supported dense/sparse cases:
  • unary, binary, scalar, reductions, matmul
• All shape/dimension validation failures return structured, user-friendly errors
• `matmul` validation honors `matmul_output_shape` contract
• Reduction semantics verified:
  • axis reductions with `keepdims`
  • full reductions (`*_all`)
  • boolean reductions (`all/any`)
• Chained transform + math correctness is validated
• Unsupported combinations return structured `Error::Unsupported(...)` (or more specific typed errors when available)
• No `todo!`, `unimplemented!`, or panic-based placeholders in production math paths

Test Plan

• Add/extend dedicated math tests in `deps/fensor/tests/`:
  • unary/binary/scalar correctness
  • reduce/reduce_all/reduce_boolean correctness
  • matmul shape and numeric correctness
  • dense/sparse parity where supported
  • transform+math composition cases
• Include negative tests for:
  • invalid shape combinations
  • incompatible matmul dimensions
  • unsupported sparse execution paths