perf: reduce per-query allocations on hot path

[adamyeats]

Summary

Reduces per-query allocations on the sqlds hot path. Every change is mechanical and preserves existing behaviour; the fleet-wide impact across ~15 Grafana SQL datasource plugins (snowflake, bigquery, clickhouse, databricks, redshift, mssql, postgres, mysql, oracle, saphana, athena, …) is proportional.

Note

Related: #240 — perf: Plumb RowCapacityHint through DBQuery to presize frame fields. The two PRs are independent and can land in either order; #240 targets per-column slice growth inside sqlutil.FrameFromRows (and is blocked on grafana/grafana-plugin-sdk-go#1536), while this PR targets allocations in sqlds itself around it. The performance wins are additive.

Five targeted changes:

• connector: cache defaultKey once on Connector instead of fmt.Sprintfing it per GetConnectionFromQuery / Connect call.
• datasource: cache driverSettings, Converters(), and the six optional mutator type assertions (QueryDataMutator, QueryMutator, ResponseMutator, QueryArgSetter, QueryErrorMutator, CheckHealthMutator) at construction; hoist settings := ds.DriverSettings() at the top of handleQuery instead of re-copying the struct 10× per query.
• query: replace applyHeaders' json.Unmarshal → map → json.Marshal round-trip with a byte-level JSON-key injection fast path; fall back to the original code on malformed input. Preallocate fixFrameForLongToMulti's time slice to the known length.
• bench_test (new): benchmark harness covering every change plus two deferred candidates (DriverSettings/Converters direct, Response_ConcurrentSet). TestMain silences backend.Logger so hclog doesn't interleave with go test output and corrupt samples.

Benchmark results

go test -bench=. -benchmem -count=10 -run=^$ ./... on Apple M2 Pro, Darwin arm64, benchstat comparison against main:

sec/op

Benchmark	Before	After	Δ
ApplyHeaders/Nil	1097.5n ± 26%	614.9n ± 22%	-43.98%
ApplyHeaders/Empty	1013.5n ± 4%	598.0n ± 2%	-41.00%
ApplyHeaders/Small	1395.0n ± 6%	621.6n ± 5%	-55.44%
ApplyHeaders/Large	3552.5n ± 25%	747.3n ± 13%	-78.96%
FixFrameForLongToMulti/100	6.389µ ± 22%	4.516µ ± 14%	-29.31%
FixFrameForLongToMulti/1000	48.80µ ± 5%	36.02µ ± 10%	-26.19%
FixFrameForLongToMulti/10000	528.9µ ± 2%	353.4µ ± 22%	-33.19%
Connector_GetConnectionFromQuery_SingleConn	118.15n ± 5%	42.63n ± 1%	-63.92%
HandleQueryMutatorChecks	4.766n ± 16%	2.076n ± 10%	-56.44%
DriverSettings	2.785n ± 16%	2.842n ± 11%	~
HandleQuery_SettingsReads	1.986n ± 2%	1.994n ± 1%	~
DriverConverters	1.975n ± 2%	1.990n ± 6%	~
DefaultKey	56.06n ± 20%	57.96n ± 13%	~
Response_ConcurrentSet/1	458.1n ± 25%	469.6n ± 9%	~
Response_ConcurrentSet/10	4.315µ ± 5%	4.229µ ± 10%	~
Response_ConcurrentSet/50	24.11µ ± 6%	26.81µ ± 14%	~
geomean	499.3n	338.9n	-32.14%

B/op

Benchmark	Before	After	Δ
ApplyHeaders/Nil	992	536	-45.97%
ApplyHeaders/Empty	992	544	-45.16%
ApplyHeaders/Small	1153	576	-50.04%
ApplyHeaders/Large	2322	776	-66.58%
FixFrameForLongToMulti/100	14.594Ki	9.867Ki	-32.39%
FixFrameForLongToMulti/1000	121.94Ki	87.96Ki	-27.86%
FixFrameForLongToMulti/10000	1588.9Ki	874.9Ki	-44.94%
Connector_GetConnectionFromQuery_SingleConn	40	0	-100%

allocs/op

Benchmark	Before	After	Δ
ApplyHeaders/Nil	17	9	-47.06%
ApplyHeaders/Empty	17	10	-41.18%
ApplyHeaders/Small	24	10	-58.33%
ApplyHeaders/Large	57	10	-82.46%
FixFrameForLongToMulti/100	122	115	-5.74%
FixFrameForLongToMulti/1000	1025	1015	-0.98%
FixFrameForLongToMulti/10000	10.03k	10.02k	-0.17%
Connector_GetConnectionFromQuery_SingleConn	2	0	-100%

Benchmarks that show no significant change (DriverSettings, HandleQuery_SettingsReads, DriverConverters, DefaultKey, Response_ConcurrentSet) were added as controls / deferred candidates — they were included in the harness before picking which fixes to land, and are kept so future regressions are detectable.

Running the benchmarks and tests

• go test ./...\
• go vet ./...\
• go test -bench=. -benchmem -count=10 -run=^$ ./...\ before and after, compare the two sets with `benchstat`