Document lk_hybrid choice in --loss-fn help text

Addresses CodeRabbit review on #673: the --loss-fn help only described
kl_div and ce; add the lk_hybrid option and point to --lk-loss-eta.

Signed-off-by: GIREESH7963 <abburigireesh@gmail.com>

Author: GIREESH7963

scripts/train.py

@@ -967,7 +967,8 @@ def parse_args():
         help=(
             "Loss function used during draft model training. "
             "'kl_div' = KL divergence (default). "
-            "'ce' = cross-entropy."
+            "'ce' = cross-entropy. "
+            "'lk_hybrid' = adaptive KL/TV blend (see --lk-loss-eta)."
         ),
     )
     parser.add_argument(