Improve JSON validation throughput by 11%

[trevorprater]

What changed

This replaces the validation string/key path with a single-pass scanner that validates control characters and escape sequences while scanning, instead of first locating strings and then doing follow-up scans/parsing.

It also avoids the strconv/strings imports on this path and keeps Validate/ValidateBytes behavior unchanged.

Why

The existing Validate benchmark fixtures spend a meaningful amount of time in string/key validation. The new scan keeps the same JSON validity checks but removes extra passes over already-scanned strings.

This patch was found while experimenting with Sleepy - https://sleepy.run/mcp, then manually reviewed and validated locally before opening this PR.

Validation

Commands run on Apple M4 / darwin arm64:

go test -count=1 ./...

Result:

ok  	github.com/valyala/fastjson	0.480s
ok  	github.com/valyala/fastjson/fastfloat	0.319s

I also ran an additional local differential test in a temporary worktree comparing both Validate and ValidateBytes against encoding/json.Valid over focused edge cases plus 20,000 deterministic randomized ASCII inputs. That check passed.

Benchmark command:

go test -run '^$' -bench '^BenchmarkValidate/(small|medium|large|canada|citm|twitter)/fastjson$' -benchmem -count=20 .

benchstat summary:

Validate/small/fastjson-10     21.86n +/- 1%  -> 18.84n +/- 1%  -13.79%
Validate/medium/fastjson-10    224.3n +/- 1%  -> 192.3n +/- 1%  -14.29%
Validate/large/fastjson-10     3.149u +/- 1%  -> 2.635u +/- 1%  -16.32%
Validate/canada/fastjson-10    222.0u +/- 0%  -> 214.9u +/- 1%   -3.19%
Validate/citm/fastjson-10      116.1u +/-11%  -> 104.5u +/- 1%  -10.01%
Validate/twitter/fastjson-10   55.94u +/- 1%  -> 49.86u +/- 2%  -10.88%
geomean                         5.304u        -> 4.693u       -11.51%

Allocations stayed at 0 B/op and 0 allocs/op.