Fix branch length optimization using correct loss; WiggleAct

[matsen]

• Fix the branch length optimization for DNSMs by using WrappedBinaryMutSel.
• Abstract DNSM classes, including a linear one
• Add the wiggle activation, and all that was required to make that work, including multiple gradient retries
• make ReduceLROnPlateau move slower
• branch length logging and restoring

Fixes #4 and #6

[matsen]

An error-tolerant wiggle:

def wiggle(x):
    output = torch.where(x < 1, 0.3 * (x - 1), 0.3 * torch.log(x))
    if torch.isnan(output).any():
        print(x)
        nan_index = torch.where(torch.isnan(output))[0][0]
        raise ValueError(f"NaN in output for input: {x[nan_index]}")
    return output

[matsen]