Improve the model we use in test and benchmark

[junpenglao]

Currently, in test and benchmark we use a simple linear regression, with the logprob function as:

def regression_logprob(scale, coefs, preds, x):
    """Linear regression"""
    scale_prior = stats.expon.logpdf(scale, 1, 1)
    coefs_prior = stats.norm.logpdf(coefs, 0, 5)
    y = jnp.dot(x, coefs)
    logpdf = stats.norm.logpdf(preds, y, scale)
    return sum(x.sum() for x in [scale_prior, coefs_prior, logpdf])

However, the free parameter scale is positive only, which can create problem for HMC. We should rewrite the logprob func so that the free parameters are in $R$ (e.g., what we are doing in https://blackjax-devs.github.io/blackjax/examples/change_of_variable_hmc.html)