docs: expand DDP metric synchronization guidance (#21685)

Restructure the "Synchronize validation and test logging" section in
accelerator_prepare.rst into a problem-framing intro plus three
subsections (sync_dist, TorchMetrics, manual all_gather), a decision
table, and a common-pitfalls list.

Directly addresses the custom-metric case: accumulate per-step outputs,
call all_gather at epoch end, and compute the metric. The "my compute
runs N times" confusion is called out and resolved — after all_gather
every rank holds the same data, so the redundant compute is cheap and
correct; only self.log needs the rank_zero_only guard.

Refs #20117

Co-authored-by: Deependu <deependujha21@gmail.com>