perf: hoist timefmt parse layouts

[sorafujitani]

Summary

• hoist timefmt.Parse layouts out of the hot-path loop
• add BenchmarkParse coverage for nano, second-precision, empty, and invalid inputs

Fixes #104

Benchmark

Command:

go test -bench=BenchmarkParse -benchmem ./internal/timefmt -count=10

benchstat /tmp/ccsession-104-before.txt /tmp/ccsession-104-after.txt:

Parse/RFC3339Nano-12   36.39n ± 2%   36.56n ± 3%       ~ (p=0.579 n=10)
Parse/RFC3339-12       23.84n ± 1%   23.56n ± 2%       ~ (p=0.183 n=10)
Parse/empty-12         1.824n ± 3%   1.808n ± 1%       ~ (p=0.796 n=10)
Parse/invalid-12       86.34n ± 1%   85.88n ± 1%       ~ (p=0.353 n=10)
geomean                19.23n        19.13n       -0.52%

Current Go already optimized the parse success paths to 0 B/op and 0 allocs/op before this change, so the measured allocation count is unchanged. The code still removes the per-call slice literal from the source-level hot path.

Verification

go test ./...
go vet ./...
golangci-lint run
gofmt -l $(git ls-files '*.go')
git diff --check