…tiguous() copy

On large prefill batches, causal_conv1d_fn receives a non-contiguous input
tensor because the GEMM output [seq, features] is transposed to [features, seq]
before being passed to the CUDA conv1d kernel. The CUDA kernel requires
stride(-1)==1, which forces a full tensor copy via .contiguous() costing
>0.6ms per layer.

The existing Triton conv1d kernel already accepts arbitrary strides by passing
stride values directly. This change falls back to the Triton path whenever the
input is non-contiguous, eliminating the copy entirely.

Tested on both embedding (Qwen3.5-0.8B) and generation workloads.

fix: avoid GPU-CPU sync when seq_lens_cpu not pre-computed

Only fall back to Triton conv1d when seq_lens_cpu is already available
in kwargs (pre-computed on CPU by the caller). When absent, keep the
original .contiguous() path to avoid introducing a GPU-CPU sync via
query_start_loc.cpu().tolist().

fix: correct fallback logic — use .contiguous() when seq_lens_cpu unavailable

Previous fix had dead code. Now clearly:
- no sgl_kernel: always use Triton (compute seq_lens_cpu if needed)
- sgl_kernel + contiguous: use CUDA kernel (fast path, unchanged)
- sgl_kernel + non-contiguous + seq_lens_cpu available: use Triton (no copy)
- sgl_kernel + non-contiguous + seq_lens_cpu absent: .contiguous() + CUDA (avoids GPU-CPU sync)

style: improve comments and docstring clarity

feat(eval): support --dataset-path for GPQA eval in run_eval.py

Allow users to pass a local CSV file path via --dataset-path instead of
downloading from the hardcoded OpenAI blob URL. Falls back to the
original URL when --dataset-path is not provided.

refactor: use use_triton variable for dispatch clarity

Address reviewer nit: unwrap the dispatch condition into a named
use_triton variable in both causal_conv1d_fn and causal_conv1d_update
for readability.

Also revert docstring format to match the original inline style,
adding a brief dispatch note at the end rather than a full rewrite.

fix: restore original docstrings, keep use_triton variable

Revert docstring changes in causal_conv1d_fn and causal_conv1d_update
to match upstream exactly. The dispatch logic explanation now lives
only in inline comments, not in the docstring.

fix: remove kwargs from causal_conv1d_update, restore clean state

- Remove **kwargs from causal_conv1d_update signature and triton call
  (no callers pass extra kwargs to the decode update function)
- Keep **kwargs only in causal_conv1d_fn where seq_lens_cpu is passed
- Docstrings now match upstream exactly