Session leak when read-write statement-based transactions are retried

[thepaul]

Environment details

• Programming language: Go
• OS: macOS, Linux
• Language runtime version: 1.23.2
• Package version: 1.7.2

Steps to reproduce

1. Use the Cloud Spanner emulator as the backend (it will abort queries more readily than the real Spanner).
2. Start a transaction tx1.
3. Execute a query (like SELECT 1) in tx1 that will cause a session to be opened. Leave that session open.
4. Outside of any transaction context, execute another single-statement query, i.e. (SELECT 2).
5. The SELECT 2 query will be aborted (because the emulator only supports one transaction at a time) and will go into an infinite retry loop using (*readWriteTransaction).retry() waiting for tx1 to close.
6. Close tx1 (rollback or commit).
7. If the SELECT 2 query has been retrying for long enough, it will have exhausted the limit on session handles and will block forever waiting for more. SELECT 2 will never complete. Also, no more sessions can be opened at this point.

[olavloite]

Fixed in #301. Further optimizations will follow once googleapis/google-cloud-go#10956 is available.