Replay native CUDA batch graphs

// Native CUDA graph replay on small seeded research weights.
// This exercises inference mechanisms, not trained language quality.
import aner.token
import aner.gen
import aner.nn

fn verify(label: String, condition: Bool) -> Unit {
    print("check_" + label)
    if !condition {
        let failure: Result<Bool, String> = Result.err(label)
        print(failure.unwrap())
    }
    print(condition)
}
fn same_ids(left: List, right: List) -> Bool {
    if left.len() != right.len() { return false }
    var index = 0
    while index < left.len() {
        if cast<Int64>(left.get(index)) != cast<Int64>(right.get(index)) { return false }
        index = index + 1
    }
    return true
}
fn same_rows(left: List, right: List) -> Bool {
    if left.len() != right.len() { return false }
    var row = 0
    while row < left.len() {
        if !same_ids(cast<List>(left.get(row)), cast<List>(right.get(row))) { return false }
        row = row + 1
    }
    return true
}

// Graph replay is explicit and requires cached CUDA execution.
// Change only Float32 to Float64 to run the other supported precision.
nn.configure("cuda", "Float32", "preserve")
let codec = token.byte_tokenizer()
let model = gen.decoder(codec, context: 96, width: 32,
    heads: 4, layers: 2, seed: 63, dropout: 0.0)
let identity = model.fingerprint()
let prompt = codec.encode("research", bos: true, eos: false)
let prompts = List(prompt, List(), List(120), prompt)
let budgets = List(24, 16, 0, 8)
let seeds = List(71, 19, 23, 97)
let graph = model.generate_batch(prompts, max_new_tokens: budgets,
    temperature: 0.8, top_k: 20, top_p: 0.95, seeds: seeds, cuda_graph: true)
let eager = model.generate_batch(prompts, budgets, 0.8, 20, 0.95, seeds, true, false)
let oracle = model.generate_batch(prompts, budgets, 0.8, 20, 0.95, seeds, false)
verify("graph_eager", same_rows(graph, eager))
verify("graph_uncached", same_rows(graph, oracle))
verify("empty_prompt_bos", same_ids(cast<List>(graph.get(1)),
    model.generate(List(), 16, 0.8, 20, 0.95, 19)))
verify("zero_budget", cast<List>(graph.get(2)).len() == 0)
verify("independent_seed", same_ids(cast<List>(graph.get(0)),
    model.generate(prompt, 24, 0.8, 20, 0.95, 71)))
let shared = model.generate_batch(List(prompt, prompt), 8, 0.0, 0, 1.0,
    List(7, 99), use_cache: true, cuda_graph: true)
verify("scalar_budget", same_ids(cast<List>(shared.get(0)), cast<List>(shared.get(1))))
let zero = model.generate_batch(prompts, List(0, 0, 0, 0), 0.0, 0, 1.0,
    seeds, cuda_graph: true)
verify("all_zero", zero.len() == 4 && cast<List>(zero.get(0)).len() == 0)
verify("model_unchanged", model.fingerprint() == identity && model.steps() == 0)
print("cuda_graph_checks_passed")