roachtest: do not fail costfuzz test on perturbed error

[michae2]

The costfuzz test repeatedly executes a randomly generated query twice,
once with a normal plan and once with a perturbed-costs plan. We were
considering a successful first execution followed by an errored second
execution a test failure. But there are certain queries that
legitimately return errors nondeterministically depending on plan. For
example:

CREATE TABLE t (a INT PRIMARY KEY);
CREATE TABLE u (b INT PRIMARY KEY);
INSERT INTO t VALUES (0);
SELECT * FROM t INNER HASH JOIN u ON a = b + 1 WHERE 1 / a = 1;
SELECT * FROM u INNER HASH JOIN t ON a = b + 1 WHERE 1 / a = 1;

The first plan will successfully return 0 rows because the hash join can
short-circuit without evaluating 1 / a. The second plan will return a
divide-by-zero error because 1 / a is evaluated while building the hash
table. This is not a bug.

Internal errors are a little more interesting than normal SQL-level
errors, so we still consider this a test failure on internal errors.

Fixes: #81032

Release note: None

[cockroach-teamcity]

This change is 

[msirek]

Reviewed 2 of 2 files at r1, all commit messages.
Reviewable status: complete! 1 of 0 LGTMs obtained (waiting on @mgartner)

[michae2]

TFTR!

bors r=msirek

[craig]

Build failed (retrying...):

• GitHub CI (Cockroach)

[craig]

Build succeeded:

• GitHub CI (Cockroach)