client: optimize destruction of WriteRpc

When writing batches with lots of operations, the WriteRpc destructor
ends up cache-miss bound, since the various InFlightOp and WriteOps are
strewn all about memory. This adds some prefetching which sped things up
noticeably (~37%) in a benchmark which ends up bound by the reactor thread on
the client side.

  $ perf stat ./build/thinlto/bin/kudu perf loadgen localhost -num_rows_per_thread=10000000 -num_threads=8

Before:
  Generator report
    time total  : 51403.6 ms
    time per row: 0.000642545 ms
  Dropping auto-created table 'default.loadgen_auto_d289807fc12a4b1c861f79b19af9ec8e'

   Performance counter stats for './build/thinlto/bin/kudu perf loadgen localhost -num_rows_per_thread=10000000 -num_threads=8':

          180,585.24 msec task-clock                #    3.508 CPUs utilized
              25,373      context-switches          #    0.141 K/sec
               1,648      cpu-migrations            #    0.009 K/sec
              50,927      page-faults               #    0.282 K/sec
     726,022,544,856      cycles                    #    4.020 GHz                      (83.33%)
      71,782,315,500      stalled-cycles-frontend   #    9.89% frontend cycles idle     (83.36%)
     412,273,652,207      stalled-cycles-backend    #   56.79% backend cycles idle      (83.29%)
     408,271,477,858      instructions              #    0.56  insn per cycle
                                                    #    1.01  stalled cycles per insn  (83.35%)
      75,750,045,948      branches                  #  419.470 M/sec                    (83.33%)
         296,247,270      branch-misses             #    0.39% of all branches          (83.34%)

        51.475433628 seconds time elapsed

       178.590913000 seconds user
         1.935099000 seconds sys

After:
  Generator report
    time total  : 37293.8 ms
    time per row: 0.000466172 ms
  Dropping auto-created table 'default.loadgen_auto_ece2f41beef94a9fa032c77899f7e61c'

   Performance counter stats for './build/thinlto/bin/kudu perf loadgen localhost -num_rows_per_thread=10000000 -num_threads=8':

          189,125.49 msec task-clock                #    5.060 CPUs utilized
              29,363      context-switches          #    0.155 K/sec
               2,043      cpu-migrations            #    0.011 K/sec
              48,405      page-faults               #    0.256 K/sec
     772,496,448,279      cycles                    #    4.085 GHz                      (83.33%)
     129,999,474,226      stalled-cycles-frontend   #   16.83% frontend cycles idle     (83.36%)
     300,049,388,250      stalled-cycles-backend    #   38.84% backend cycles idle      (83.30%)
     414,415,517,571      instructions              #    0.54  insn per cycle
                                                    #    0.72  stalled cycles per insn  (83.32%)
      76,829,647,882      branches                  #  406.236 M/sec                    (83.34%)
         352,749,453      branch-misses             #    0.46% of all branches          (83.35%)

        37.376785122 seconds time elapsed

       186.834651000 seconds user
         2.143945000 seconds sys

Change-Id: I538f995f7ec161e746885c6b31cd1dccd72139b0
Reviewed-on: http://gerrit.cloudera.org:8080/14868
Reviewed-by: Adar Dembo <adar@cloudera.com>
Tested-by: Todd Lipcon <todd@apache.org>
2 files changed