Add a mean-log-likelihood method to improve the ACT estimation

Suggestion from @colm.talbot. I've then added a number of other tweaks in addition.

bilby/core/sampler/ptemcee.py

@@ -618,14 +618,17 @@ class Ptemcee(MCMCSampler):
 
 def get_mean_logl_min_it(mean_log_likelihood, frac):
     nsteps = mean_log_likelihood.shape[1]
-    if nsteps > 10:
-        zero_chain_mean_log_likelihood = mean_log_likelihood[0, :]
-        maxl = np.max(zero_chain_mean_log_likelihood)
-        fracdiff = (maxl - zero_chain_mean_log_likelihood) / np.abs(maxl)
+    if nsteps < 10:
+        return 0
+
+    min_it = 0
+    for x in mean_log_likelihood:
+        maxl = np.max(x)
+        fracdiff = (maxl - x) / np.abs(maxl)
         idxs = fracdiff < frac
         if np.sum(idxs) > 0:
-            return np.min(np.arange(len(idxs))[idxs])
-    return 0
+            min_it = np.max([min_it, np.min(np.arange(len(idxs))[idxs])])
+    return min_it
 
 
 def check_iteration(
@@ -994,9 +997,10 @@ def plot_tau(
 def plot_mean_log_likelihood(mean_log_likelihood, outdir, label):
 
     ntemps, nsteps = mean_log_likelihood.shape
-    max_logl = np.max(mean_log_likelihood)
-    ymax = max_logl + 0.01 * np.abs(max_logl)
-    ymin = max_logl - 0.1 * np.abs(max_logl)
+    ymax = np.max(mean_log_likelihood)
+    ymin = np.min(mean_log_likelihood[:, -100:])
+    ymax += 0.1 * (ymax - ymin)
+    ymin -= 0.1 * (ymax - ymin)
 
     fig, ax = plt.subplots()
     idxs = np.arange(nsteps)