GLM-5.3-Flash on main: export fails and prefix-cache restores fail with device error 72 #272

GLM-5.3-Flash can't be served from main (0572fd1). There are two independent problems.

1. Export fails while compiling, on every rank:

torch._dynamo.exc.UserError: Could not extract specialized integer from data-dependent expression u51
  ... block_fp8.py _project -> plan_from_handle(handle)

Since 4a24d78, _project is traced and looks up its plan from a handle that preparation marks as unbacked.

2. Every prefix-cache restore fails with prepared model reported device error 72 once export works. The first request succeeds and every later request fails.

Merge 2b38bf9 kept #165's two-page trailing regions but restored the older trailing_groups[region.nbytes] key (state_transfer.py:273). That key is two pages, so each shifted MTP draft copy writes two pages at the wrong offset. The copy overwrites last_hidden and consumed_tokens, and the next restore fails validation. Only GLM-5.3 passes trailing pages, so no other model is affected.

Repro (2x RTX PRO 6000, SM120):

ignition export glm-5.3-flash --gpus 0,1 --max-concurrency 8 --prefill-chunk-size 4096 \
  --mtp-draft-tokens 5 --speculative-token-capacity 48 --language-model-only

Serve the core and send two chat requests. A core exported at f7c468c, before the merge, serves normally.