Add experimental JavaScript backend and Node slice-sampling PoC

[ricardoV94]

Just playing with it

def sample_normal_mean(*, draws=10_000, tune=1_000, seed=12345):
    observed = np.array([1.2, 0.8, 1.1, 0.9], dtype="float64")
    with pm.Model() as model:
        pm.Normal("mu", mu=0.0, sigma=1.0)
        pm.Normal("y", mu=model["mu"], sigma=1.0, observed=observed)

    fn = pytensor.function(model.value_vars, model.logp(), mode="JS")
    try:
        program = fn.vm.jit_fn.program
        assert len(program.inputs) == len(program.outputs) == 1
        payload = {
            "program": {
                "source": program.source,
                "constants": [
                    {
                        "shape": list(value.shape),
                        "data": base64.b64encode(value.tobytes()).decode("ascii"),
                    }
                    for value in program.constants
                ],
                "inputs": [
                    {"ndim": variable.ndim, "dtype": variable.type.dtype}
                    for variable in program.inputs
                ],
                "outputs": [
                    {"ndim": variable.ndim, "dtype": variable.type.dtype}
                    for variable in program.outputs
                ],
            },
            "initial": float(model.initial_point()["mu"]),
            "draws": draws,
            "tune": tune,
            "width": 1.0,
            "maxSteps": 20,
            "seed": seed,
        }
    finally:
        fn.vm.jit_fn.close()

    completed = subprocess.run(
        ["node", str(Path(__file__).with_name("slice_sampler.mjs"))],
        input=json.dumps(payload),
        text=True,
        capture_output=True,
        check=True,
    )
    result = json.loads(completed.stdout)
    result["posterior_mean"] = float(observed.sum() / (1 + len(observed)))
    result["posterior_sd"] = float(1 / np.sqrt(1 + len(observed)))
    return result