Add token-id LM training, CUDA defaults, and kernel launch hygiene

[wadkisson]

Summary

Three library changes

Part A — Token-id causal LM API

• Add causalTransformerTokenLmScalarModuleDef for float-encoded token-id inputs
• Add text helpers: causalLmTokenRows, causalLmTokenFloatVec, causalLmTokenSampleRowsFromTokenArray
• Add floatVecToNatTensor op (Functional → Ops → TorchLean Functional → Trainer instances)
• Add verifier stubs in Compile.lean and SpecEval.lean

Enables persistent-module LM training: token windows can change each step without rebuilding the module or one-hot tensors (Adam state stays on one session). Matches PyTorch nn.Embedding + cross_entropy structurally.

Part B — CUDA CLI + inference

• --cuda auto-enables --fast-kernels (fastKernels := fastKernels || useGpu)
• Add Options.noGrad; eval1NoGrad sets noGrad := true
• README: CUDA quickstart no longer requires explicit --fast-kernels

Makes --cuda turn on fast kernels by default and fixes inference so eval1NoGrad runs without recording autograd state.

Part C — Native CUDA robustness

• Add torchlean_cuda_clear_pending_error and torchlean_cuda_check_launch
• Use before scalar unary kernel launches and in reduceMean

Clears stale CUDA driver errors before kernel launches so unrelated failures don’t break TorchLean’s native kernels.