moe(w4a16): preserve preplanned tile geometry across launch op

[voipmonitor]

Problem

compile_w4a16_fused_moe(..., force_tile_config=...) correctly builds a pinned launch, but run_w4a16_moe crosses the torch.library custom-op boundary and _w4a16_fused_moe_launch_flat recompiles without the selected tile geometry. The lookup therefore silently resolves auto tiles. NF3 packs are tile-N specific, so the GLM-5.2 TP4 hybrid path packed FC1 for tile-N 256 while the custom op ran the auto tile-N 128 kernel, producing garbled output.

The production-shape reproducer measured relative error 1.29 with the pinned (64,256,64,256) launch before this fix. Small auto-tile tests did not expose it.

Fix

Thread the four selected FC1/FC2 tile dimensions through both normal and calibrated launch custom ops, then re-pin those exact values at the internal compile/cache lookup. This preserves the preplanned launch contract without disabling direct-topk TC decode or duplicating weights.

Validation

• New GLM-5.2 TP4 production-shape NF3 test: H=6144, I/rank=512, m=8 launch replayed at m=1, pinned tiles, top-k=8, and -1 cross-tier routes.
• test_w4a16_moe_launch_ops_have_fake_dispatch passes with the extended schemas.
• NF3 numerical suite plus preplanned TC-decode suite: 11 passed.
• Full madeby561/GLM-5.2-MXFP8-NVFP4-NF3-Hybrid TP4 E2E with the native vLLM loader: short and 30k-context generation are coherent; the same image without this patch is garbled from token 1.