fix: do not delete existing object on upload (#5778)

This increases performance for concurrent writes for azure and gcs
backends as they had the check inside lock. For all backends this
reduces the number of roundtrips to backend needed. For the blob
backends the writes atomically overwrites any existing blob. So this
should be faster and safer, there is otherwise a short window of time
where no blob exists which could lead to unexpected results for reads.

This PR have have been tested with
#5779, see it for example and
timings.

 ---
TYPE: IMPROVEMENT
DESC: Reduce latency of object storage uploads

Author: efredriksson-modelon

tiledb/sm/filesystem/azure.cc

@@ -1115,11 +1115,6 @@ void Azure::write_blocks(
 
     auto state_iter = block_list_upload_states_.find(uri.to_string());
     if (state_iter == block_list_upload_states_.end()) {
-      // Delete file if it exists (overwrite).
-      if (is_file(uri)) {
-        remove_file(uri);
-      }
-
       // Instantiate the new state.
       BlockListUploadState state;
 

tiledb/sm/filesystem/gcs.cc

@@ -1034,11 +1034,6 @@ Status GCS::write_parts(
 
       // Downgrade to read lock, expected below outside the create
       unique_wl.unlock();
-
-      // Delete file if it exists (overwrite).
-      if (is_file(uri)) {
-        remove_file(uri);
-      }
     } else {
       // If another thread switched state, switch back to a read lock
       state = &multi_part_upload_states_.at(uri.to_string());

tiledb/sm/filesystem/s3.cc

@@ -1227,10 +1227,6 @@ void S3::global_order_write(
     state_ptr = &state_iter->second;
 
     unique_wl.unlock();
-
-    if (is_file(uri)) {
-      remove_file(uri);
-    }
   } else {
     state_ptr = &state_iter->second;
     // Unlock, as make_upload_part_req will reaquire as necessary.