Make integration tests more tolerant of transient failures

Our integration tests often perform "non-idempotent" operations, but these can fail even if our code is working fine. A typical test may have this structure:

```cc
auto create = client.CreateFoo(....);
ASSERT_STATUS_OK(create);

auto update = client.UpdateFoo(...);
ASSERT_STATUS_OK(update);

auto del = client.DeleteFoo(...);
ASSERT_STATUS_OK(del);
```

None of those operations are retried by default, as the operations are non-idempotent. I think we could rewirte the tests as follows:

```cc
auto create = client.CreateFoo(... Options{}.set(RetryEverything()));
ASSERT_THAT(create, StatusIs(AnyOf(StatusCode::kOk, StatusCode::kAlreadyExists)));
```

Or if we don't want to retry the operation, we can write it as:

```cc
auto create = client.CreateFoo(...);
if (IsTransient(...)) GTEST_SKIP() << create.status();
```

This probably requires a mini-design doc, and will require breaking down the work per test.

For now I am going to consolidate all existing "transient error on non-idempotent operation" flakes to this one.

The issue names no files or specific tests; begin by reviewing the existing transient-error-on-non-idempotent-operation flakes consolidated here. Draft the requested mini-design, then break the work down per integration test, deciding whether each operation should retry or skip on transient failure; completion criteria are not specified beyond making these tests more tolerant.

索引モデルが issue の本文から書いたものです。