Log the skipped optimizer step on gradient overflow (#8417)
ZeRO returns early from `step()` on non-finite gradients.
`DynamicLossScaler` reports that from inside `update_scale`, but
`CreateLossScaler` only returns it for `fp16 + dynamic_scaling` — bf16
and fp32 always get the static `LossScaler`, whose `update_scale` is an
inherited no-op.
The skip is counted in `engine.skipped_steps`, but that is only surfaced
by `_report_progress`, gated on `steps_per_print`, and
`STEPS_PER_PRINT_DEFAULT` is `None`. So on a default bf16 config the
loop advances, the LR scheduler steps and metrics are emitted while no
weights are updated, with no output at all.
Fixes #8415
---------
Signed-off-by: Abhishek Kulkarni <abhishek.kulkarni@snowflake.com>
Co-authored-by: Olatunji Ruwase <tunji.ruwase@snowflake.com>
