test: add benchmarks

[olavloite]

Adds benchmarks that compare the performance of the go-sql driver with the normal Spanner client.

Executing the benchmarks on a VM on Google Cloud that is physically close to the database that is being tested yields the following results:

Benchmark	go-sql	Client
SelectSingleRecord	8027885 ns/op	7605408 ns/op
SelectAndUpdateUsingMutation	16588398 ns/op	15454486 ns/op
SelectAndUpdateUsingDml	20143182 ns/op	21006488 ns/op
CreateAndReload	16834755 ns/op	16987523 ns/op
CreateAlbumsUsingMutations.	22605913 ns/op	23259244 ns/op
Select100Singers	92737263 ns/op	94163536 ns/op
Select100SingersInReadOnlyTx	96486050 ns/op	96695674 ns/op
Select100SingersInReadWriteTx	130437345 ns/op	124886457 ns/op

Closes #33

[hengfengli]

@olavloite Does this mean that the performance of go-sql driver is pretty similar to directly using client library? They look very close and there is no big difference.

[hengfengli]

LGTM.

[olavloite]

@olavloite Does this mean that the performance of go-sql driver is pretty similar to directly using client library? They look very close and there is no big difference.

Yes, that is correct. The Go sql driver is in most cases a relatively thin layer on top of the client library, so that is also as expected. The only scenario where there is a measurable difference is when selecting a larger result set in a read/write transaction, as the Go sql driver will then keep track of a running checksum of the results that it sees.

[rahul2393]

Fresh run

Benchmark	go-sql	Client
SelectSingleRecord	7256458 ns/op	6365998 ns/op
SelectAndUpdateUsingMutation	35894568 ns/op	30139076 ns/op
SelectAndUpdateUsingDml	32422615 ns/op	21006488 ns/op
CreateAndReload	33583545 ns/op	24802121 ns/op
CreateAlbumsUsingMutations.	40948656 ns/op	45372519 ns/op
Select100Singers	42971542 ns/op	35908125 ns/op
Select100SingersInReadOnlyTx	39085144 ns/op	49537134 ns/op
Select100SingersInReadWriteTx	58363907 ns/op	45240976 ns/op