Kafka read performance regression due to added header support

Support for headers in KafkaIO reads was recently added:
https://issues.apache.org/jira/browse/BEAM-10865

This introduced several reflection calls into the path of advancing KafkaUnboundedReader. While separately running benchmarks, I noticed this regression.

Calls currently come from:
ConsumerSpEL.hasHeaders -> can be cached similar to other booleans
deserialize key and value methods -> could be avoided in cases where headers are not being examined (at a minimum can be avoided for known coders like ByteArrayDeserializer)