Address review of non-reentrant activation checkpoint CPU offload
Keep Wing Lian's original commits; this follow-up restores the
GradientCheckpointingLayer marker on context exit, skips offloading
the last checkpoint input, and clarifies bytes vs count knobs.
Signed-off-by: Olatunji Ruwase <tunji.ruwase@snowflake.com>
Co-authored-by: Cursor <cursoragent@cursor.com>