Pin FlashInfer with serialized TP4 BF16 two-shot reductions

Use published FlashInfer revision 03de0b663abe8f5dc69d4e10118587ab459b13eb
for preparation and record the same revision in the lockfile. This enables
single-rounding BF16 two-shot dispatch for serialized TP4 messages in the
64–640 KiB band while retaining caller-owned IPC storage.

Validation: dependency resolution and installation pass. The rebased source
passes 798 Helios tests (18 ignored), 186 focused Python tests (3 skipped),
and 15 CUDA tests on GPUs 4–7, including TP4 reductions, readiness timing,
compiled attention metadata, and native expert-sum replay. The release CLI
and Heliopause builds pass. Full-model throughput was not rerun after rebase.

Design-Signed-Off-By: Codex (GPT-6); design-review=6aeb

Author: lukealonso

pyproject.toml

@@ -22,7 +22,7 @@ prepare = [
     "torch==2.14.1+cu132",
     "torchvision==0.29.1+cu132",
     "transformers[torch,vision,chat_template] @ git+https://github.com/huggingface/transformers.git@1d9654e78a23fb167ef539cbd699dfbd3e4830cf",
-    "flashinfer-python[b12x] @ git+https://github.com/local-inference-lab/flashinfer.git@622cec10eb75da189f4dadddabc2d142a33cd211",
+    "flashinfer-python[b12x] @ git+https://github.com/local-inference-lab/flashinfer.git@03de0b663abe8f5dc69d4e10118587ab459b13eb",
     "cuda-python==13.3.1",
     "safetensors==0.8.0",
     "tokenizers==0.23.2",