Honor max_fwd_stash_count=0 in activation offload
A zero forward stash still retained one in-flight GPU activation because the
source tensor was stashed unconditionally after reaping, so the documented
max_fwd_stash_count + keep_last_count peak understated real GPU usage. The
record_stream call already keeps the allocation alive for the duration of the
D2H copy, so the Python reference can be dropped immediately.
Signed-off-by: Olatunji Ruwase <tunji.ruwase@snowflake.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
