fix: handle initial and dtype in reductions

[Robertboy18]

Summary

• cast reduction inputs when dtype is specified
• use the caller-provided initial value instead of always using the identity
• validate tensor and array initializers are scalar
• add coverage for sum, prod, max, and min initial values and explicit reduction dtypes

Problem

The HLO reduction helper accepted initial and dtype but ignored both. As a result, np.sum(x, initial=10) still initialized the reduction with zero, and np.sum(x, dtype=np.float16) still returned float32. The resulting HLO compiled successfully but had different semantics from NumPy.

Testing

• pytest -q tests/unit/test_tensor_api.py -k "reduction or prod" (30 passed)
• compared emitted HLO values, shapes, and dtypes with NumPy across 48 cases
• compiled the regression kernels for trn2 with NeuronX Compiler 2.26

I did not have Trainium hardware available, so validation covered HLO semantics and compilation rather than device execution.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[ymwangg]

neuronx-cc currently doesn't support init value in reduce op, opened a ticket to compiler team NCC-9303.