feat(sm70): support Volta / Tesla V100 via torch 2.10 cu128 downgrade

[itongxiaojun]

Motivation

Torch 2.11 — the only torch line the stock dependency pins allow — dropped
sm_70 (Volta) kernels in every CUDA build, so FreeToken cannot run at all on
Tesla V100 / other Volta GPUs: no kernel image is available for execution on the device on every CUDA op, before even reaching the runtime kernels.
This branch unblocks Volta with a dependency-level change only — no engine
code changes required, since the runtime's capability gates and pure-Triton
fallback already cover sm_70.

Changes

• [build-system] and dependencies: torch floor lowered from
  >=2.11,<2.12 to >=2.10,<2.11. torch 2.10.0+cu128 is the last 2.x line
  that ships sm_70 cubins and runs on driver r570 (CUDA 12.8) — no driver
  upgrade or reboot needed on existing Volta nodes.
• [tool.uv.index]: pytorch-cu130 → pytorch-cu128
  (https://download.pytorch.org/whl/cu128) so uv resolves the +cu128 build.
• Extras: flashinfer-python[cu13] → [cu12]; sglang-kernel dropped
  from accel (no cu128 wheel for 0.4.5). On Volta both are unusable anyway
  (sm_80+/sm_100+ only) and the runtime falls back to pure-Triton kernels.
• Docs: sm_70 install path, Desktop-engine caveats (uninstall
  sglang/flashinfer + leftover nvidia/cu13 libs on in-place cu130→cu128
  upgrades), and 16 GiB V100 serving tuning for large system prompts.

Verification (Tesla V100-SXM2-16GB, driver 570.211.01, CUDA 12.9 toolkit)

• uv pip install -e . source build with nvcc 12.9 — OK
• torch.cuda.get_arch_list() includes sm_70; matmul on V100 — OK
• ft checkpoint: Qwen3.6-35B-A3B-FP8 (35 GB) → FTW conversion — OK
• ft serve: OpenAI-compatible /v1/* chat with reasoning + streaming
  (progressive chunks + [DONE]) — OK; ~20 tok/s decode
• Wired into an OpenAI-compatible agent harness (DSH llm-pi-ai) with the
  tuned serve flags in the docs.

Usage

git clone -b feat/support-sm_70 https://github.com/redoop/FreeToken.git
export PATH=/usr/local/cuda-12.9/bin:$PATH CUDA_HOME=/usr/local/cuda-12.9
uv venv --python 3.12 .venv-sm70
uv pip install -e .   # base package — no [accel] on Volta