Fix decoder cache publication race

[tirathawat]

Summary

• make decoder fast-cache publication atomic to avoid torn Decoder interface reads under concurrent first use
• share the same atomic cache path between race and non-race decoder builds
• add regression coverage for issue enconter panic when calling unmarshal() #495 and internal cache winner behavior

Test Plan

• go test -count=1 ./...
• go test -race -count=1 ./...
• go test -count=100 ./internal/decoder -run 'TestDecoderFastCache'
• go test -count=100 . -run TestIssue495

Fixes #495

[tirathawat]

Hi @goccy, this PR fixes #495 by making decoder cache publication atomic. Could you take a look when you have time?