add direct_write_embedding method

Summary:
X-link: pytorch/FBGEMM#4800

X-link: facebookresearch/FBGEMM#1824

This method will write the input embedding directly without through optimizer

Reviewed By: q10

Differential Revision: D79290124

Author: emlin

torchrec/distributed/batched_embedding_kernel.py

@@ -321,6 +321,11 @@ def _populate_zero_collision_tbe_params(
         enable_optimizer_offloading=True,
         backend_return_whole_row=(backend_type == BackendType.DRAM),
         eviction_policy=eviction_policy,
+        embedding_cache_mode=(
+            config.fused_params.get("embedding_cache_mode", False)
+            if config.fused_params
+            else False
+        ),
     )