[YSQL] Use lightweight protobuf for writing to tserver

[lnguyen-yugabyte]

Description

Currently when running bulk insert / copy command, in the postgres backend for, about 15 percent of CPU time is spent on memory allocation / deallocation.

     6.12%  postgres         libtcmalloc.so.4.5.6         [.] operator new[]
     5.42%  postgres         postgres                     [.] CopyReadLine
     2.48%  postgres         postgres                     [.] NextCopyFromRawFields
     2.46%  postgres         libtcmalloc.so.4.5.6         [.] operator delete[]
     2.32%  postgres         libyb_pggate.so              [.] yb::pggate::PgTableDesc::GetColumnInfo
     2.25%  postgres         libyb_pggate.so              [.] yb::pggate::PgApiImpl::NewConstant
     2.05%  postgres         postgres                     [.] nocachegetattr
     1.88%  postgres         libc-2.23.so                 [.] __memcpy_avx_unaligned
     1.78%  postgres         libtcmalloc.so.4.5.6         [.] tcmalloc::CentralFreeList::FetchFromOneSpansSafe
     1.48%  postgres         postgres                     [.] pg_verify_mbstr_len
     1.37%  postgres         libyb_common.so              [.] yb::QLType::Create
     1.35%  postgres         postgres                     [.] GetTablePrimaryKeyBms
     1.21%  rpc_tp_pggate_y  libtcmalloc.so.4.5.6         [.] operator delete[]
     1.11%  postgres         postgres                     [.] YBCExecuteInsertInternal
     1.10%  postgres         libyb_common_proto.so        [.] yb::QLValuePB::ByteSizeLong
     1.08%  rpc_tp_pggate_y  libtcmalloc.so.4.5.6         [.] tcmalloc::CentralFreeList::ReleaseToSpans
     1.03%  postgres         libyb_pggate.so              [.] yb::pggate::PgExpr::internal_type
     1.00%  postgres         libyb_pggate.so              [.] boost::unordered::detail::table<boost::unordered::detail::map<std::__1::allocator<std::__1::pair<yb::PgsqlExpressionPB* const, yb::pggate::PgExpr*> >, yb::PgsqlExpressionPB*, yb::pggate::PgExpr*, boost::hash<yb::PgsqlExpressionPB*>, std::__1::equal_to<yb::PgsqlExpressionPB*> > >::try_emplace_unique<yb::PgsqlExpressionPB* const&>
     0.98%  postgres         libtcmalloc.so.4.5.6         [.] tcmalloc::CentralFreeList::RemoveRange
     0.98%  postgres         libyb_common_proto.so        [.] yb::PgsqlExpressionPB::ByteSizeLong
     0.95%  pggate_ybclient  libcrypto.so.1.1             [.] _aesni_ctr32_ghash_6x
     0.92%  rpc_tp_pggate_y  libyb_common_proto.so        [.] yb::PgsqlWriteRequestPB::~PgsqlWriteRequestPB
     0.92%  postgres         libyb_common_proto.so        [.] yb::QLValuePB::clear_value

Hence we could see good reduction of copy time if we optimize for this allocation/deallocation process. One of such idea is to use Protobuf's arena: https://developers.google.com/protocol-buffers/docs/reference/arenas. This will optimize for allocating & deallocation the protobufs itself.

We'd like to make this change for the copy/write path to the tserver and test how much time reduction it would take with this change enabled.

[lnguyen-yugabyte]

Initial performance test (on copying 10M rows):

Baseline: 5:27.483
Lightweight protobuf: 5:05.505

(e.g. copy time reduced by 6.8% compared to baseline).

[lnguyen-yugabyte]

Arena on PGConstant: 4:43.265 (13.5% time reduction compared to baseline).