5080deploy: merge upstream #198, reserve explicit KV pages during MoE auto-sizing

[gdevenyi]

Stacked on #53. Merges upstream FlashML-org#198 (head 271420a) into 5080deploy, rebuilt on upstream main cac247a (v0.1.3).

Why it is in scope for this deployment: Generic planner fix (--num-tokens with --moe-cache-auto).

Merge: engine.py merged clean. Conflict only in test_engine_resolve_auto_moe_cache_size_maps_kwargs: the PR rewrites it around a monkeypatched resolve_moe_cache_auto; kept that form and re-expressed fork #20's slot-limit check as captured["max_slots"] == 5.

Testing (RTX 5080 16 GB, production flags: FTW-fp8, nvfp4 KV, 262,144 tokens, 1,280 expert slots):

• tests/engine/test_cache_budget.py: 30 passed.
• pytest tests/ -m "not slow" at 69a5efa (the tip before the middleware commit 8adde91, which touches only server/api_server.py and adds one test; tests/server rerun at 8adde91: 871 passed): 2412 passed, 206 skipped; the 4 failures are environment-only (3x test_muse_glimmer_vision.py fail identically on pristine upstream cac247a with transformers 5.15.1, and test_async_pull_coalesced is an async def test that needs pytest-asyncio, which the project does not declare).

Rebased 2026-09-19 onto upstream main cc1f5c2 (4 commits past cac247a: FlashML-org#471 greedy sampling in mixed batches, FlashML-org#518 WeightLoadError, FlashML-org#521 tvm-ffi jit arch, FlashML-org#524 install index). Same replay as before via git rerere; the replayed stack differs from the previous tip 8adde91 by exactly the cac247a..cc1f5c2 file set. This PR's head is now a663d9e.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Bu6LgoxLR4wETqb7RPR2vt