fix(cuda/quantized): chunk long-prompt qmm launches to avoid gridDim>65535 and int overflow

[inureyes]

Summary

Long-prompt prefill of 4-bit models aborted on the CUDA backend with an invalid launch (cuGraphAddKernelNode / cuLaunchKernelEx "invalid argument", SIGABRT). This surfaced as FAIL:bench cells in the long-prompt sweep (epic #623/#624) and as crashes in mlxcel generate on long prompts. Two distinct causes in the quantized-matmul path:

• MoE GatherQMM: qmm_sm80 sets grid.z = l = tokens * num_experts_per_tok, which exceeds the CUDA gridDim.z limit of 65535 once tokens*top_k >= 65536 (qwen3-30b-a3b @8192, mixtral-8x7b @32768).
• Dense LM head: l = out.size() / (m * n) is computed with int32 m,n, which overflows when tokens*vocab >= 2^31 (about 16744 tokens for a 128256 vocab), yielding l = 0 and an invalid grid.z of 0.

Fix

All in the mlxcel-owned overlay src/lib/mlx-cpp/patches/mlx/backend/cuda/quantized/quantized.cpp:

• run_batch_chunked splits GatherQMM's gathered batch into <= 65535 slices (sub-array views over the same buffer).
• run_row_chunked splits QuantizedMatmul's M into slices of min(65535, INT32_MAX/N) so each launch keeps grid.y <= 65535 and m*n < 2^31. Gated on single_batch = out.size() == size_t(M)*N (the LM-head output is 3D [1, tokens, vocab], so an ndim==2 gate would miss it).
• int B = out.size() / M / N (sequential 64-bit division) so the overflow no longer yields B=0 and misroutes dispatch.

Both helpers early-return to the original single call when no split is needed, so normal-size matmuls are byte-identical. The change is CUDA-only; the Metal backend already uses out.size()/M/N and has no 65535 grid cap, so it is unaffected.

Validation (GB10, NVIDIA sm_121)

Previously-failing long-prompt cells now pass, exit 0 with real throughput:

• Bench (what scripts/bench_longprompt.sh runs): llama-3.1-8b-4bit @17000 / @32768, qwen3-30b-a3b-4bit @8192.
• mlxcel generate: llama-3.1-8b-4bit @23491 and qwen3-30b-a3b-4bit @~11k complete with coherent output, confirming the sub-array offsets are correct.

Closes #648