[ashhart]

Part of #1657: Kolibri-1 on CUDA keeps its prompts in the shared paged KV store, as Gemma 4 does since #1883.

What changes

• With TF_KV_SHARE=N (off by default, at most N kept states), Kolibri-1's kept prompt states live in the core KV pool's pages instead of private copies. A resumed prompt restores its page-aligned kept pages and its chronological ring rows, and gives the same tokens as a cold one.
• The store's accounting matches Gemma 4's term for term: blob pages times page bytes, copy bytes, and the layout with no snapshot bytes.
• The narrower 128-dim paged prompt kernels load only when Kolibri opts in, and in the kernel contract. Otherwise a missing kernel returns WindowShape. Spark's page loading is unchanged from Gemma 4 on CUDA: prompts in the shared KV store with TF_KV_SHARE (part of #1657) #1883.

Unset or 0 keeps today's copies, byte for byte.

Checks

• Host: a real-loader host test for the page load, plus the store tests. The lean check shows 424 problems before and after.
• Two independent reviews: LAND.
• GB10, Kolibri-1, defaults (main against this change, balanced order): GREEN. Prompt tokens are identical in all 20 rows, and decode at 1 to 8 streams is within noise (8 streams: 293.1 against 293.0 tok/s).
• GB10 with TF_KV_SHARE=8 against main, to check that replies stay equal with the store on: results will be posted here before this lands.

[ashhart]

Closing in favour of @jschmied's #1755 (with #1620), which did the same Kolibri-1 store wiring first. See the comparison and the plan on #1755.