rpc,server: investigate regression caused by #138368

[tbg]

Branch here.

We observed that #138368 reliably lowers performance by ~1% as measured by time/op in BenchmarkSysbench/SQL/3node/oltp_read_write.

This commit (server: give Node.BatchStream its own impl) not only claws back this 1% regression, it actually achieves another 1% improvement on top (when comparing "pr plus my commit" vs "master before pr"). Remarkably, the code change in this commit is a partial revert of a refactor introduced in the PR, and since nothing in the PR should've made performance better (after all, the PR introduced another setting to check, and the setting is off), it's surprising that the commit actually gains 1.73% vs master:

old:  f9df57e Merge #137984
new:  826b922 server: give Node.BatchStream its own impl
args: benchdiff "--old" "lastmerge" "./pkg/sql/tests" "-b" "-r" "Sysbench/SQL/3node/oltp_read_write" "-d" "1000x" "-c" "20"
name                                   old time/op    new time/op    delta
Sysbench/SQL/3node/oltp_read_write-24    11.7ms ± 1%    11.5ms ± 1%  -1.73%  (p=0.000 n=19+20)

name                                   old alloc/op   new alloc/op   delta
Sysbench/SQL/3node/oltp_read_write-24    2.20MB ± 3%    2.19MB ± 3%    ~     (p=0.968 n=20+20)

name                                   old allocs/op  new allocs/op  delta
Sysbench/SQL/3node/oltp_read_write-24     10.9k ± 2%     10.9k ± 2%    ~     (p=0.703 n=20+20)

To circle in on this more, I wanted to see if pr+commit vs pr produces a 2% speed-up. I didn't even have to run this for very long, it's pretty clear that it does (p-value is basically zero):

old:  b0ea57c server: fix drpc end-to-end test
new:  826b922 server: give Node.BatchStream its own impl
args: benchdiff "--old" "HEAD~" "./pkg/sql/tests" "-b" "-r" "Sysbench/SQL/3node/oltp_read_write" "-d" "1000x" "-c" "20"

running benchmarks:
 name                                   old time/op    new time/op    delta
Sysbench/SQL/3node/oltp_read_write-24    11.9ms ± 1%    11.6ms ± 1%  -2.33%  (p=0.000 n=8+8)

name                                   old alloc/op   new alloc/op   delta
Sysbench/SQL/3node/oltp_read_write-24    2.17MB ± 4%    2.19MB ± 2%    ~     (p=0.959 n=8+8)

name                                   old allocs/op  new allocs/op  delta
Sysbench/SQL/3node/oltp_read_write-24     10.8k ± 2%     10.8k ± 2%    ~     (p=0.627 n=8+8)

I don't know what to make of this yet, but am reminded of this segment of "Performance Matters" by Emery Berger.

I'm nearing the end of my day, but leaving this here as food for thought for now.

Closes #139022.

[blathers-crl]

It looks like your PR touches production code but doesn't add or edit any test code. Did you consider adding tests to your PR?

_{🦉 Hoot! I am a Blathers, a bot for CockroachDB. My owner is dev-inf.}

[cockroach-teamcity]

This change is 

[RaduBerinde]

LGTM

I was not able to reproduce the gains with a similar change, but this looks better regardless.

I'd refactor the other caller of batchStreamImpl as well since there's only one, no need for all the interface stuff