PERF: `broadcast_apply_full_axis` can be implemented to compute caches in some cases

[anmyachev]

[anmyachev]

not actual

[YarShev]

I think we could compute caches not only when keep_partitioning=False, but also when keep_partitioining=True so I would prefer to keep this issue open.

Source comment: #5640 (comment)