Convert ops_common.py to re-export shell (#5794)

Summary:
Pull Request resolved: #5794

Convert split_table_batched_embeddings_ops_common.py from a 528-line
module defining all TBE common types to a ~60-line re-export shell that
imports from the canonical locations:

- Core types (EmbeddingLocation, PoolingMode, etc.) from fbgemm_gpu.tbe.config
- Cache types (CacheAlgorithm, CacheState) from fbgemm_gpu.tbe.cache.cache_config
- SSD types (EvictionPolicy, KVZCHParams, etc.) from fbgemm_gpu.tbe.ssd.ssd_config

Infrastructure changes:
- Create lightweight :tbe_cache_config and :tbe_ssd_config BUCK targets
- Remove cache_config.py/ssd_config.py from their parent ops targets' srcs
- Add get_new_embedding_location() to tbe/config/embedding_config.py

All existing import paths continue to work unchanged.

Reviewed By: henrylhtsang

Differential Revision: D103477971

fbshipit-source-id: cf18816bdc8d42e0d839595b584058854de64760

Author: q10

fbgemm_gpu/fbgemm_gpu/sparse_ops.py

@@ -14,7 +14,7 @@
 
 # fmt:skip
 from fbgemm_gpu.split_embedding_configs import SparseType
-from fbgemm_gpu.split_table_batched_embeddings_ops_common import PoolingMode
+from fbgemm_gpu.tbe.config import PoolingMode
 from fbgemm_gpu.utils.loader import load_torch_module
 
 try:

fbgemm_gpu/fbgemm_gpu/split_embedding_configs.py

@@ -14,10 +14,7 @@
 import torch
 
 # fmt:skip
-from fbgemm_gpu.split_table_batched_embeddings_ops_common import (
-    EmbeddingLocation,
-    SplitState,
-)
+from fbgemm_gpu.tbe.config import EmbeddingLocation, SplitState
 
 
 def pad4(value: int) -> int:

fbgemm_gpu/fbgemm_gpu/split_embedding_inference_converter.py

@@ -20,14 +20,14 @@
     QuantizationConfig,
     SparseType,
 )
-from fbgemm_gpu.split_table_batched_embeddings_ops_common import EmbeddingLocation
 from fbgemm_gpu.split_table_batched_embeddings_ops_inference import (
     IntNBitTableBatchedEmbeddingBagsCodegen,
 )
 from fbgemm_gpu.split_table_batched_embeddings_ops_training import (
     ComputeDevice,
     SplitTableBatchedEmbeddingBagsCodegen,
 )
+from fbgemm_gpu.tbe.config import EmbeddingLocation
 from fbgemm_gpu.tbe.utils import quantize_embs
 from torch import Tensor  # usort:skip
 

fbgemm_gpu/fbgemm_gpu/split_table_batched_embeddings_ops.py

@@ -18,17 +18,6 @@
 #   fbgemm_gpu.split_table_batched_embeddings_ops_training
 
 from fbgemm_gpu.split_embedding_configs import EmbOptimType as OptimType, SparseType
-from fbgemm_gpu.split_table_batched_embeddings_ops_common import (
-    BoundsCheckMode,
-    CacheAlgorithm,
-    CacheState,
-    DEFAULT_SCALE_BIAS_SIZE_IN_BYTES,
-    EmbeddingLocation,
-    PoolingMode,
-    RecordCacheMetrics,
-    round_up,
-    SplitState,
-)
 from fbgemm_gpu.split_table_batched_embeddings_ops_inference import (
     align_to_cacheline,
     IntNBitTableBatchedEmbeddingBagsCodegen,
@@ -48,6 +37,16 @@
     TailIdThreshold,
     WeightDecayMode,
 )
+from fbgemm_gpu.tbe.cache import CacheAlgorithm, CacheState
+from fbgemm_gpu.tbe.config import (
+    BoundsCheckMode,
+    DEFAULT_SCALE_BIAS_SIZE_IN_BYTES,
+    EmbeddingLocation,
+    PoolingMode,
+    RecordCacheMetrics,
+    round_up,
+    SplitState,
+)
 
 try:
     if torch.version.hip: