store: writers serialise instead of failing under contention !235

Stacked on !234. Last of the v1.12.0 six.

Every transaction in the store package writes. A deferred one takes the write lock at its first write, by which point another writer may hold it; SQLite answers SQLITE_BUSY and does not run the busy handler for that case, so busy_timeout cannot help and the transaction fails outright.

Measured, eight concurrent writers and eight readers over two seconds:

44% of concurrent read-then-write transactions fail today. immediate takes the lock at BEGIN, where busy_timeout applies, and the failures go to zero — readers gain too, because failing writers currently spin.

SetMaxOpenConns(1), the issue's other suggestion, is the third row: it also removes the failures but caps write throughput for reader throughput this instance does not need.

TestConcurrentWritersDoNotFail fails 229 of 320 without the DSN change; I verified that by reverting it.

Behaviour change worth knowing: a heavily contended write now blocks for up to busy_timeout (5s) instead of returning an error immediately.

Closes #121