fix(moe_vec): chunk launches when tokens*top_k exceeds gridDim.z limit - #125

Problem

The MMVQ MoE kernels in csrc/gguf/moe_vec.cuh place tokens * top_k in the
z dimension of the launch grid:

const dim3 block_nums(block_num_y, 1, tokens * top_k);

gridDim.z is hardware-capped at 65535. vLLM's default
max_num_batched_tokens is 8192, so any MMVQ-path MoE model with top_k >= 8
hits 8192 * 8 = 65536 > 65535 inside profile_run and the engine dies
during startup with:

RuntimeError: CUDA error: invalid argument

This was found while integrating the Hy3 (Hunyuan) architecture (#88), whose
IQ1_M experts (top_k 8) hit this limit in profile_run on every startup.

This is 100% reproducible on startup for such models with default settings —
it is not a load-dependent or intermittent failure. Decode-time batches are
small, so the limit is only reachable through the profile/dummy run or large
prefill batches.

The vec path is selected when both expert weight types are in
MMVQ_QUANT_TYPES but not on the MMQ gemm path
(quantization/fused_moe.py), which in practice means i-matrix (IQ*) quants —
the formats commonly used for large MoE GGUFs.

Fix

Split the launch into per-token-range chunks so each launch keeps
gridDim.z <= 65535, mirroring the existing chunking pattern in
quantize_row_q8_1_cuda (gguf_kernel.cu):

const int max_tokens_per_launch = 65535 / top_k;
for (int tok_off = 0; tok_off < tokens; tok_off += max_tokens_per_launch) {
  ...
}

Chunk boundaries align to whole tokens (multiples of top_k in grid z), and
the vy / dst / topk_ids pointers are offset accordingly, so kernel
indexing is unchanged. Single-launch cases take the same path as before with
one iteration.

The 19 per-type launcher bodies were byte-identical except for template
arguments, so they are consolidated into one MOE_VEC_Q8_1_LAUNCHER macro
(429 → 158 lines) while touching the launch logic.

Test Plan / Test Result

New regression test test_moe_large_batch exercises the chunked launch path
directly: num_tokens=9000, top_k=8 gives 9000 * 8 = 72000 > 65535, which
crashed with CUDA error: invalid argument before the fix. It runs for both
IQ1_M (the motivating i-matrix case) and Q4_0, and compares against the
dequantized fused_experts reference.

Full run on an RTX PRO 6000 Blackwell (sm_120), torch 2.13.0+cu130, CUDA 13
toolchain, vLLM 0.27.0:

pytest tests/test_kernels.py::test_moe_large_batch -v
2 passed in 287.04s

pytest tests/test_kernels.py
1452 passed, 68 failed, 576 skipped in 3338.46s

pytest tests/test_plugin.py tests/test_gguf_utils.py
44 passed in 8.62s

The 68 failures are all test_moe[*-dtype2-...] cases failing inside the
triton reference kernel with OutOfResources: shared memory, Required: 122880, Hardware limit: 101376 — a hardware/triton limitation of this GPU unrelated
to the change. The identical failure set (same test IDs, same error) exists in
the baseline run on the parent commit recorded before this branch
(69 failed there = the same 68 kernel cases plus one unrelated
test_plugin.py ordering flake that passes in isolation and passes here).

ruff check on the full repo also passes.