Replaced APPROX_ADAGRAD and APPROX_ROWWISE_ADAGRAD with EXACT_ADAGRAD…

… and EXACT_ROWWISE_ADAGRAD

Summary:
Updated `APPROX_ADAGRAD` and `APPROX_ROWWISE_ADAGRAD` to `EXACT_ADAGRAD` and `EXACT_ROWWISE_ADAGRAD`.

Removed `to_exact` function and usage.

Removed `APPROX_` usage in enums.

Fixed test with `EXACT_SGD` instead of `SGD`

Reviewed By: jianyuh

Differential Revision: D26380060

fbshipit-source-id: 1ea806414944df296a53ae604b03b39814ca5415

fbgemm_gpu/split_embedding_configs.py

@@ -15,10 +15,6 @@ class EmbOptimType(enum.Enum):
     )
     LAMB = "lamb"
     ADAM = "adam"
-    # approx refers to that gradients to the same row are applied in sequence
-    # instead of first coalescing gradients then apply together
-    APPROX_ADAGRAD = "adagrad"
-    APPROX_ROWWISE_ADAGRAD = "approx_row_wise_adagrad"
     # exact/dedup: gradients to the same row are applied with coalesce then apply
     # together, instead of applied in sequence (approx).
     EXACT_ADAGRAD = "exact_adagrad"