self.log: `strategy.reduce` vs `reduce_fx`

Description & Motivation

self.log used to have separate parameters sync_dist_op and reduce_fx: The reduce_fx callable was used to reduce all of the logged values over an entire epoch, whereas sync_dist_op was used to reduce values across multiple GPUs. Currently it is not possible to set sync_dist_op. Instead, strategy.reduce is used but there is no way to pass the optional reduce_op parameter to this function. I think this means you always get the default, which is 'mean' for DDPStrategy. I think it would be useful for self.log to expose both the between-gpu reduction and the over-epoch reductions.
For example, if one wants to log the maximum value encountered on any GPU over an entire epoch, one would call self.log(..., on_epoch=True, sync_dist=True, sync_dist_op='max', reduce_fx='max'). self.log can then create a partial(strategy.reduce, reduce_op=sync_dist_op) so that the strategy reduce is used, but with the desired reduce_op.
Today, if you do self.log(..., on_epoch=True, sync_dist=True, reduce_fx='max') you will get the maximum over the epoch of the mean value across all GPUs, which is not what is desired, and could be confusing.

Pitch

self.log (and log_dict) should take a sync_dist_op parameter and use it as the reduce_op kwarg for strategy.reduce to improve logging values over multiple GPUs and multiple steps per epoch. The sync_dist_op and reduce_fx parameters should be renamed to make it clear why they are different and what they do.

Alternatives

Use a torchmetrics.Metric instead.

Additional context

No response

cc @borda @carmocca @Blaizzy

Start by tracing the self.log and log_dict entry points and the strategy.reduce call, focusing on how reduce_op and reduce_fx currently interact. Compare the current multi-GPU and epoch aggregation behavior with the max-over-GPUs example in the issue. Done means the API clearly supports separate cross-device and over-epoch reductions without the current default-mean ambiguity.