Measure the warmup transient before CUDA graph capture
Capture and later setup add persistent allocations, so a peak - current
read at sizing time under-measured the warmup steps' transient (by 0.31 GiB
for Gemma4 MTP on MI355, enough with fragmentation to hit the torch cap).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Xgy3B2NXY4GnQaw2WwsqpC
Signed-off-by: Nick Hill <nickhill123@gmail.com>
