GLM CUDA: keep kept prompt states in pinned host memory (complements #1732) #1735

GLM CUDA keeps its kept prompt states in one device arena, which gets only the memory the KV page pool leaves over
(#1510). On a discrete card that can be a fraction of a GiB, so a long conversation's state is refused and every turn
prefills the whole conversation again. #1732 addresses this by keeping prompts as the pool's own pages.

This issue is a complementary route for discrete cards: keep kept states in pinned host memory.

• The server already has a host-cache path (host_snaps, --prompt-cache-gib sized in host memory, as Nemotron uses);
  GLM doesn't opt in.
• With the snapshot arena in pinned host memory mapped into the device's address space, the existing
  cuMemcpyDtoDAsync keep/restore copies move a state across PCIe (~57 GB/s each way on PCIe Gen5 x16).
• The device cache's share goes back to the KV pool.
• Unset, nothing changes.

Where it helps most is many long conversations at once (agent sessions):

• glm cuda: TF_GLM_KV_SHARE keeps prompts as the pool's pages in prompt_radix's store #1732's pool pages hold one hot conversation at device speed.
• Host memory holds many: tens of GiB instead of the pool's few GiB.
• The two combine: with glm cuda: TF_GLM_KV_SHARE keeps prompts as the pool's pages in prompt_radix's store #1732 on, its pool-page states' fixed parts move to host memory too.

Measured on two RTX PRO 6000 Blackwell cards (96 GB) at TP=2 with Mia-AiLab/GLM-5.3-Flash-EXL3-4bpw-TensorFold
(weights 86.2 GiB a rank, 320K window):

Resumed replies are identical to fresh ones, and bench_concurrent --alone --serial passes. A pull request with the
full receipt follows.

This is the first of a few steps toward serving GLM-5.3-Flash on two 96 GB cards with system memory taking the pressure
off the KV window. Later steps would be GLM's vision tower in host memory, then live DSA latent rows in host memory for
long windows, building on the host tier in #1501/#1522. Each would be its own issue and pull request, and we'd like
maintainer guidance on the last one before starting it.