Add OPSD example: config, divergence losses, utils + tests
First slice of the on-policy distillation example app under examples/opsd/.
This commit lands the framework-agnostic foundation: the OPSDConfig dataclass
hierarchy, chunked / streamed forward-KL / reverse-KL / JSD losses with
sequence-axis chunking to bound peak memory, response-mask + shift helpers,
and a 24-case CPU-only test suite covering identity, masking, chunk
equivalence, gradient flow, and numerical edge cases.
Signed-off-by: Zhipeng Wang <zhipengbayern@gmail.com>
