[query] Move LoweredTableReaderCoercer into ExecuteContext

[ehigham]

Refactored table reader coercion and caching mechanism.

What changed?

• Removed shouldCacheQueryInfo method from Backend class
• Introduced CoercerCache in ExecuteContext
• Refactored LoweredTableReader.makeCoercer to return a function instead of a class
• Removed local caching in GenericTableValue and LoweredTableReader

Why make this change?

This change aims to optimize table reader coercion by:

• Centralizing caching logic in ExecuteContext
• Allowing more flexible caching strategies across different backend implementations

[ehigham]

• [query] Move is.hail.check to test code #14801 : 2 dependent PRs (#14693 , #14842 )
• [query] Move LoweredTableReaderCoercer into ExecuteContext #14696 👈 (View in Graphite)
• [query] Remove persistedIr from Backend interface #14692
• [query] Remove lookupOrCompileCachedFunction from Backend interface #14691
• [query] Remove BlockMatrix persist from Backend interface #14690
• [query] Expose references via ExecuteContext #14686
• [query] Extract Backend Methods called from Python into Py4JBackendExtensions #14767
• main

This stack of pull requests is managed by Graphite. Learn more about stacking.

[patrick-schultz]

Nice change. While I was looking into what LoweredTableReaderCoercer does, I noticed it's only used in the VCF and Plink readers. That makes me think we could get rid of LoweredTableReaderCoercer and GenericTableValue completely if we refactored those readers to create TableStages directly, probably using registered functions to do the actual parsing logic. But thats just something to think about for the future.