--moe-cache-auto asserts "cache budget too small" instead of falling back to prefill overlap off #642

plan_cache_budget (engine/cache_budget.py:72-74) turns on prefill overlap, and with it a 2*num_experts slot floor, whenever the slot cap allows it, without checking budget_bytes. When the budget fits num_experts slots plus the KV reserve but not 2*num_experts, the plan is clamped up to the floor, and the total <= budget_bytes assert (line 89) fires instead of the overlap-off fallback at line 81. This happens after the full expert-bank load, and the error doesn't mention --disable-moe-prefill-overlap, which would fit.

from freetoken.engine.cache_budget import plan_cache_budget
a = (2557061534, 2764800, 1683456, 512, 24576)
plan_cache_budget(*a, True, 128, 24576)   # AssertionError: cache budget too small (moe=1024 slots)
plan_cache_budget(*a, False, 128, 24576)  # (846, 129, False)

Expected: fall back to overlap off, and log it, when 2*num_experts slots plus the KV reserve don't fit.

Environment: main 07813248 (source build), RTX 5080 16 GB (sm_120), driver 610.57.04, Ubuntu 24.04, Python 3.14, torch 2.11.0+cu130, triton 3.6.0; checkpoint nvidia/Qwen3.8-Flash-Next-NVFP4.

Implementation (on my fork, with a test): 27fe671. When the slot floor does not fit, the plan drops prefill overlap instead of asserting. Fork PR: gdevenyi#20