Move all step-based GS functionality to nodes (#94)

Summary:

X-link: facebook/Ax#4746

X-link: facebook/Ax#4731

A lot going on here:

1. Make `GenerationStep` a factory for `GenerationNode` by replacing its `__init__(self, ...)` constructor (which would have to return a `GStep`) with a `__new__(cls, ...)` construtor, which can return a `GNode` (magic, thanks Devmate!)
2. Adapt storage: stop storing steps and just treat them as nodes. No step-only fields will be saved going forward. Backward compatibility is handled though.
3. Change a bazillion tests and checks in downstream applications.

NOTE: need to remember next steps here: https://www.internalfb.com/diff/D86066476?dst_version_fbid=732904879835890&transaction_fbid=1400151778204976, cc mgarrard 

sorrybigdiff

Reviewed By: mgarrard

Differential Revision: D80128678

Author: Lena Kashtelyan

tests/test_tuning.py

@@ -344,9 +344,18 @@ def test_warm_starting_trials_produces_the_right_number_of_sobol_and_bayesian_tr
         )
 
     value_counter = trial_results["generation_node"].value_counts().to_dict()
-    sobol_count = value_counter["GenerationStep_0"]
-    botorch_count = value_counter["GenerationStep_1"]
-
+    sobol_key = (
+        "GenerationStep_0_Sobol"
+        if "GenerationStep_0_Sobol" in value_counter
+        else "GenerationStep_0"
+    )
+    sobol_count = value_counter[sobol_key]
+    botorch_key = (
+        "GenerationStep_1_BoTorch"
+        if "GenerationStep_1_BoTorch" in value_counter
+        else "GenerationStep_1"
+    )
+    botorch_count = value_counter[botorch_key]
     expected_botorch = total_trials - n_warmup_random_trials - 1
     assert (
         len(trial_results) == total_trials